Spaces:
Running
Running
File size: 112,510 Bytes
1e83cc1 efd2c7f 1e83cc1 019892c 1e83cc1 5a55f05 1e83cc1 ff22b96 1e83cc1 5a55f05 1e83cc1 efd2c7f 1e83cc1 efd2c7f 1e83cc1 ff927bf 1e83cc1 ff22b96 1e83cc1 52b2e03 1e83cc1 d6a7310 52b2e03 decc7d3 1e83cc1 decc7d3 1e83cc1 decc7d3 1e83cc1 efd2c7f 1e83cc1 d6a7310 1e83cc1 decc7d3 1e83cc1 efd2c7f d6a7310 1e83cc1 d6a7310 1e83cc1 d6a7310 1e83cc1 d6a7310 1e83cc1 efd2c7f 1e83cc1 d6a7310 1e83cc1 d6a7310 1e83cc1 d6a7310 1e83cc1 d6a7310 1e83cc1 decc7d3 1e83cc1 d6a7310 1e83cc1 d6a7310 1e83cc1 d6a7310 1e83cc1 d6a7310 1e83cc1 d6a7310 1e83cc1 d6a7310 1e83cc1 d6a7310 1e83cc1 019892c 1e83cc1 decc7d3 1e83cc1 d6a7310 52b2e03 d6a7310 1e83cc1 d6a7310 1e83cc1 decc7d3 1e83cc1 efd2c7f decc7d3 efd2c7f decc7d3 1e83cc1 decc7d3 1e83cc1 decc7d3 1e83cc1 decc7d3 1e83cc1 decc7d3 1e83cc1 decc7d3 1e83cc1 decc7d3 1e83cc1 efd2c7f ff22b96 5a55f05 decc7d3 1e83cc1 decc7d3 1e83cc1 ff22b96 1e83cc1 ff22b96 1e83cc1 ff22b96 1e83cc1 ff22b96 1e83cc1 ff22b96 1e83cc1 ff22b96 5a55f05 1e83cc1 decc7d3 1e83cc1 efd2c7f 1e83cc1 efd2c7f 1e83cc1 5a55f05 ff22b96 5a55f05 1e83cc1 efd2c7f 1e83cc1 efd2c7f 1e83cc1 52b2e03 d6a7310 52b2e03 d6a7310 1e83cc1 52b2e03 1e83cc1 decc7d3 1e83cc1 decc7d3 1e83cc1 decc7d3 1e83cc1 decc7d3 1e83cc1 decc7d3 1e83cc1 decc7d3 1e83cc1 ff927bf 1e83cc1 019892c 1e83cc1 019892c 1e83cc1 019892c 1e83cc1 019892c 1e83cc1 019892c 1e83cc1 019892c 1e83cc1 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 | <!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Cham-OCR: A Deep-Learning OCR Pipeline for Unicode Eastern Cham, toward Historical Manuscript and Epigraphic Recognition — Technical Report</title>
<!-- Google Fonts: Google Sans Flex & JetBrains Mono -->
<link rel="preconnect" href="https://fonts.googleapis.com">
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
<link href="https://fonts.googleapis.com/css2?family=Google+Sans+Flex:wght@400;500;600;700&family=JetBrains+Mono:wght@400;500;600&family=Noto+Sans+Cham:wght@400;600;700&display=swap" rel="stylesheet">
<!-- KaTeX for pristine mathematical typesetting -->
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/katex@0.16.8/dist/katex.min.css">
<script defer src="https://cdn.jsdelivr.net/npm/katex@0.16.8/dist/katex.min.js"></script>
<script defer src="https://cdn.jsdelivr.net/npm/katex@0.16.8/dist/contrib/auto-render.min.js" onload="renderMathInElement(document.body);"></script>
<style>
:root {
--bg-page: #FBF8F3;
--bg-surface: #FFFFFF;
--bg-surface-elevated: #F5EFE6;
--bg-subtle: #F2EBE0;
--border-color: #E8DFD1;
--border-strong: #D5C8B4;
--text-primary: #2D2926;
--text-secondary: #6B6257;
--text-muted: #8F8577;
--accent-primary: #C96442;
--accent-hover: #B25232;
--accent-subtle: #F7EAE3;
--accent-green: #2E7D32;
--accent-green-bg: #E8F5E9;
--accent-blue: #1565C0;
--accent-blue-bg: #E3F2FD;
--font-sans: 'Google Sans Flex', 'Google Sans', system-ui, -apple-system, sans-serif;
--font-mono: 'JetBrains Mono', monospace;
--font-cham: 'Noto Sans Cham', 'Google Sans Flex', sans-serif;
--shadow-sm: 0 1px 3px rgba(45, 41, 38, 0.05);
--shadow-md: 0 4px 12px rgba(45, 41, 38, 0.08);
--radius-sm: 6px;
--radius-md: 10px;
--radius-lg: 14px;
}
* {
box-sizing: border-box;
margin: 0;
padding: 0;
}
html {
scroll-behavior: smooth;
scroll-padding-top: 85px;
}
section[id], h2[id], h3[id], [id] {
scroll-margin-top: 85px;
}
body {
font-family: var(--font-sans);
background-color: var(--bg-page);
color: var(--text-primary);
line-height: 1.72;
font-size: 15.5px;
-webkit-font-smoothing: antialiased;
}
/* Top Header Navigation Bar */
.top-navbar {
position: sticky;
top: 0;
z-index: 100;
background: rgba(251, 248, 243, 0.95);
backdrop-filter: blur(10px);
border-bottom: 1px solid var(--border-color);
padding: 12px 24px;
display: flex;
justify-content: space-between;
align-items: center;
}
.nav-brand {
display: flex;
align-items: center;
gap: 10px;
text-decoration: none;
color: var(--text-primary);
font-weight: 700;
font-size: 16px;
}
.nav-brand svg {
color: var(--accent-primary);
}
.nav-actions {
display: flex;
gap: 12px;
align-items: center;
}
.nav-btn {
display: inline-flex;
align-items: center;
gap: 6px;
padding: 6px 14px;
font-size: 13.5px;
font-weight: 600;
border-radius: var(--radius-sm);
text-decoration: none;
transition: all 0.15s ease;
cursor: pointer;
border: 1px solid var(--border-color);
background: var(--bg-surface);
color: var(--text-primary);
}
.nav-btn:hover {
background: var(--bg-subtle);
border-color: var(--border-strong);
color: var(--accent-primary);
}
.nav-btn-primary {
background: var(--accent-primary);
border-color: var(--accent-primary);
color: #FFFFFF;
}
.nav-btn-primary:hover {
background: var(--accent-hover);
color: #FFFFFF;
}
/* Page Layout */
.page-container {
max-width: 1280px;
margin: 0 auto;
padding: 40px 24px 80px;
display: grid;
grid-template-columns: 280px 1fr;
gap: 48px;
align-items: start;
}
/* Sidebar Navigation */
.sidebar {
position: sticky;
top: 76px;
max-height: calc(100vh - 96px);
overflow-y: auto;
padding: 16px;
background: var(--bg-surface);
border: 1px solid var(--border-color);
border-radius: var(--radius-md);
box-shadow: var(--shadow-sm);
}
.toc-title {
font-size: 12px;
font-weight: 700;
text-transform: uppercase;
letter-spacing: 0.8px;
color: var(--text-muted);
margin-bottom: 12px;
padding-bottom: 8px;
border-bottom: 1px solid var(--border-color);
}
.toc-list {
list-style: none;
}
.toc-item {
margin-bottom: 4px;
}
.toc-link {
display: block;
padding: 6px 10px;
font-size: 13px;
color: var(--text-secondary);
text-decoration: none;
border-radius: var(--radius-sm);
transition: all 0.15s ease;
line-height: 1.4;
}
.toc-link:hover {
background: var(--bg-subtle);
color: var(--accent-primary);
}
.toc-link.active {
background: var(--accent-subtle);
color: var(--accent-primary);
font-weight: 600;
}
.toc-sublist {
list-style: none;
padding-left: 14px;
margin-top: 2px;
}
.toc-sublink {
display: block;
padding: 4px 8px;
font-size: 12px;
color: var(--text-muted);
text-decoration: none;
border-radius: var(--radius-sm);
transition: all 0.15s ease;
}
.toc-sublink:hover {
color: var(--accent-primary);
}
/* Article Main Content */
.article-main {
background: var(--bg-surface);
padding: 48px 56px;
border: 1px solid var(--border-color);
border-radius: var(--radius-lg);
box-shadow: var(--shadow-sm);
min-width: 0;
}
/* Header / Meta */
.paper-tag-strip {
display: flex;
gap: 8px;
flex-wrap: wrap;
margin-bottom: 16px;
}
.tag-pill {
font-size: 11.5px;
font-weight: 600;
padding: 3px 10px;
border-radius: 20px;
text-transform: uppercase;
letter-spacing: 0.5px;
}
.tag-pill-accent {
background: var(--accent-subtle);
color: var(--accent-primary);
}
.tag-pill-green {
background: var(--accent-green-bg);
color: var(--accent-green);
}
.tag-pill-blue {
background: var(--accent-blue-bg);
color: var(--accent-blue);
}
.paper-title {
font-size: 30px;
font-weight: 700;
line-height: 1.25;
color: var(--text-primary);
margin-bottom: 18px;
letter-spacing: -0.4px;
}
.author-box {
padding: 16px 20px;
background: var(--bg-page);
border: 1px solid var(--border-color);
border-radius: var(--radius-md);
margin-bottom: 28px;
display: flex;
justify-content: space-between;
align-items: center;
flex-wrap: wrap;
gap: 12px;
}
.author-info {
display: flex;
flex-direction: column;
gap: 2px;
}
.author-name {
font-size: 16px;
font-weight: 700;
color: var(--text-primary);
}
.author-affiliation {
font-size: 13.5px;
color: var(--text-secondary);
}
.author-meta {
display: flex;
gap: 16px;
font-size: 12.5px;
color: var(--text-muted);
}
/* Abstract Box */
.abstract-box {
padding: 22px 26px;
background: var(--bg-subtle);
border-left: 4px solid var(--accent-primary);
border-radius: 0 var(--radius-md) var(--radius-md) 0;
margin-bottom: 36px;
}
.abstract-title {
font-size: 13px;
font-weight: 700;
text-transform: uppercase;
letter-spacing: 0.8px;
color: var(--accent-primary);
margin-bottom: 8px;
}
.abstract-text {
font-size: 14.5px;
color: var(--text-primary);
line-height: 1.68;
text-align: justify;
}
.keywords-strip {
margin-top: 12px;
font-size: 13px;
color: var(--text-secondary);
}
.keywords-strip strong {
color: var(--text-primary);
}
/* Section Headings */
h2 {
font-size: 22px;
font-weight: 700;
color: var(--text-primary);
margin: 44px 0 16px;
padding-bottom: 8px;
border-bottom: 1.5px solid var(--border-color);
letter-spacing: -0.3px;
display: flex;
align-items: center;
gap: 8px;
}
h3 {
font-size: 17px;
font-weight: 600;
color: var(--accent-primary);
margin: 28px 0 12px;
}
h4 {
font-size: 15px;
font-weight: 600;
color: var(--text-primary);
margin: 20px 0 8px;
}
p {
margin-bottom: 16px;
color: var(--text-primary);
text-align: justify;
}
ul, ol {
margin: 0 0 18px 24px;
color: var(--text-primary);
}
li {
margin-bottom: 6px;
}
/* Cham Font Inline / Blocks */
.cham-script {
font-family: var(--font-cham);
font-size: 1.15em;
color: var(--accent-primary);
background: var(--bg-page);
padding: 1px 6px;
border-radius: 4px;
border: 1px solid var(--border-color);
display: inline-block;
}
/* Scientific Callouts & Alerts */
.academic-callout {
padding: 16px 20px;
margin: 20px 0;
border-radius: var(--radius-md);
border: 1px solid var(--border-color);
background: var(--bg-page);
}
.callout-title {
font-weight: 700;
font-size: 13.5px;
margin-bottom: 6px;
display: flex;
align-items: center;
gap: 6px;
}
.callout-info {
background: #F4F8FC;
border-color: #D2E3F7;
color: #0C4A6E;
}
.callout-info .callout-title { color: #0284C7; }
.callout-warning {
background: #FEF7EE;
border-color: #F8DFC0;
color: #7C2D12;
}
.callout-warning .callout-title { color: #EA580C; }
.callout-success {
background: #F0FDF4;
border-color: #BBF7D0;
color: #14532D;
}
.callout-success .callout-title { color: #16A34A; }
/* Tables */
.table-responsive {
overflow-x: auto;
margin: 22px 0;
border: 1px solid var(--border-color);
border-radius: var(--radius-md);
box-shadow: var(--shadow-sm);
}
table {
width: 100%;
border-collapse: collapse;
font-size: 13.5px;
text-align: left;
background: var(--bg-surface);
}
th {
background: var(--bg-subtle);
color: var(--text-primary);
font-weight: 700;
padding: 10px 14px;
border-bottom: 2px solid var(--border-strong);
white-space: nowrap;
}
td {
padding: 10px 14px;
border-bottom: 1px solid var(--border-color);
color: var(--text-primary);
}
tr:nth-child(even) td {
background: rgba(251, 248, 243, 0.4);
}
tr:hover td {
background: var(--bg-subtle);
}
.num {
font-family: var(--font-mono);
font-size: 13px;
text-align: right;
}
th.num {
text-align: right;
}
.metric-badge {
font-family: var(--font-mono);
font-weight: 600;
padding: 2px 7px;
border-radius: 4px;
font-size: 12px;
}
.metric-good {
background: var(--accent-green-bg);
color: var(--accent-green);
}
.metric-warn {
background: #FEF3C7;
color: #92400E;
}
.metric-bad {
background: #FEE2E2;
color: #991B1B;
}
/* Mathematical Display */
.math-block {
margin: 18px 0;
padding: 14px 20px;
background: var(--bg-page);
border: 1px solid var(--border-color);
border-radius: var(--radius-md);
overflow-x: auto;
text-align: center;
}
/* Code Blocks */
.code-box {
font-family: var(--font-mono);
font-size: 13px;
background: #23201D;
color: #E6E1DC;
padding: 16px 20px;
border-radius: var(--radius-md);
overflow-x: auto;
margin: 18px 0;
line-height: 1.6;
border: 1px solid #3A3530;
position: relative;
}
.code-box pre {
margin: 0;
font-family: inherit;
}
.code-header {
display: flex;
justify-content: space-between;
align-items: center;
margin-bottom: 8px;
padding-bottom: 6px;
border-bottom: 1px solid #3A3530;
font-size: 11.5px;
color: #9C948B;
text-transform: uppercase;
letter-spacing: 0.5px;
}
.copy-btn {
background: #3A3530;
color: #E6E1DC;
border: none;
padding: 3px 8px;
border-radius: 4px;
font-size: 11px;
cursor: pointer;
transition: all 0.15s;
}
.copy-btn:hover {
background: var(--accent-primary);
color: #FFFFFF;
}
/* References Section */
.ref-item {
margin-bottom: 12px;
padding-left: 28px;
text-indent: -28px;
font-size: 13.5px;
color: var(--text-secondary);
line-height: 1.6;
}
.ref-item a {
color: var(--accent-primary);
text-decoration: none;
word-break: break-all;
}
.ref-item a:hover {
text-decoration: underline;
}
/* Footer */
.paper-footer {
margin-top: 60px;
padding-top: 24px;
border-top: 1px solid var(--border-color);
display: flex;
justify-content: space-between;
align-items: center;
flex-wrap: wrap;
gap: 16px;
font-size: 13px;
color: var(--text-muted);
}
/* Responsive */
@media (max-width: 960px) {
.page-container {
grid-template-columns: 1fr;
gap: 24px;
padding: 20px 16px;
}
.sidebar {
display: none;
}
.article-main {
padding: 24px 20px;
}
.paper-title {
font-size: 24px;
}
}
@media print {
.top-navbar, .sidebar, .nav-actions, .copy-btn {
display: none !important;
}
.page-container {
grid-template-columns: 1fr;
padding: 0;
}
.article-main {
border: none;
box-shadow: none;
padding: 0;
}
body {
background: #FFF;
color: #000;
font-size: 12pt;
}
}
</style>
</head>
<body>
<!-- Top Sticky Header -->
<header class="top-navbar">
<a href="/" class="nav-brand">
<svg width="22" height="22" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.2" stroke-linecap="round" stroke-linejoin="round">
<path d="M14 2H6a2 2 0 0 0-2 2v16a2 2 0 0 0 2 2h12a2 2 0 0 0 2-2V8z"></path>
<polyline points="14 2 14 8 20 8"></polyline>
<line x1="16" y1="13" x2="8" y2="13"></line>
<line x1="16" y1="17" x2="8" y2="17"></line>
<polyline points="10 9 9 9 8 9"></polyline>
</svg>
<span>Cham OCR Technical Report</span>
</a>
<div class="nav-actions">
<a href="/" class="nav-btn">
<svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><path d="M19 12H5M12 19l-7-7 7-7"/></svg>
<span>Back to OCR Studio</span>
</a>
<a href="https://github.com/phucsd/Cham-OCR" target="_blank" rel="noopener noreferrer" class="nav-btn">
<svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><path d="M9 19c-5 1.5-5-2.5-7-3m14 6v-3.87a3.37 3.37 0 0 0-.94-2.61c3.14-.35 6.44-1.54 6.44-7A5.44 5.44 0 0 0 20 4.77 5.07 5.07 0 0 0 19.91 1S18.73.65 16 2.48a13.38 13.38 0 0 0-7 0C6.27.65 5.09 1 5.09 1A5.07 5.07 0 0 0 5 4.77a5.44 5.44 0 0 0-1.5 3.78c0 5.42 3.3 6.61 6.44 7A3.37 3.37 0 0 0 9 18.13V22"></path></svg>
<span>GitHub Repo</span>
</a>
<button class="nav-btn nav-btn-primary" onclick="window.print()">
<svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"><polyline points="6 9 6 2 18 2 18 9"></polyline><path d="M6 18H4a2 2 0 0 1-2-2v-5a2 2 0 0 1 2-2h16a2 2 0 0 1 2 2v5a2 2 0 0 1-2 2h-2"></path><rect x="6" y="14" width="12" height="8"></rect></svg>
<span>Print / Save PDF</span>
</button>
</div>
</header>
<div class="page-container">
<!-- Sidebar Navigation -->
<aside class="sidebar">
<div class="toc-title">Table of Contents</div>
<ul class="toc-list">
<li class="toc-item"><a href="#abstract" class="toc-link active">Abstract</a></li>
<li class="toc-item">
<a href="#sec1" class="toc-link">1. Introduction & Paleography</a>
<ul class="toc-sublist">
<li><a href="#sec1-1" class="toc-sublink">1.1 Historical Script Typology</a></li>
<li><a href="#sec1-2" class="toc-sublink">1.2 Abugida Orthography & Glyphs</a></li>
<li><a href="#sec1-3" class="toc-sublink">1.3 Logical vs Visual Unicode</a></li>
</ul>
</li>
<li class="toc-item">
<a href="#sec2" class="toc-link">2. Related Work</a>
<ul class="toc-sublist">
<li><a href="#sec2-1" class="toc-sublink">2.1 Epigraphic Glyph Recognition</a></li>
<li><a href="#sec2-2" class="toc-sublink">2.2 Deep Learning & Transliteration</a></li>
<li><a href="#sec2-3" class="toc-sublink">2.3 Manuscript Preservation</a></li>
<li><a href="#sec2-4" class="toc-sublink">2.4 System Positioning</a></li>
</ul>
</li>
<li class="toc-item">
<a href="#sec3" class="toc-link">3. Detection & Line Segmentation</a>
<ul class="toc-sublist">
<li><a href="#sec3-1" class="toc-sublink">3.1 Diacritic Collision Problem</a></li>
<li><a href="#sec3-2" class="toc-sublink">3.2 Cham-DBNet Architecture</a></li>
<li><a href="#sec3-3" class="toc-sublink">3.3 Indic Valley-Cut Heuristics</a></li>
</ul>
</li>
<li class="toc-item">
<a href="#sec4" class="toc-link">4. Recognition Architecture</a>
<ul class="toc-sublist">
<li><a href="#sec4-1" class="toc-sublink">4.1 SVTR-LCNet Backbone</a></li>
<li><a href="#sec4-2" class="toc-sublink">4.2 Joint CTC & Attention Loss</a></li>
<li><a href="#sec4-3" class="toc-sublink">4.3 Canonical Normalization</a></li>
</ul>
</li>
<li class="toc-item">
<a href="#sec5" class="toc-link">5. Synthetic Generation Pipeline</a>
<ul class="toc-sublist">
<li><a href="#sec5-1" class="toc-sublink">5.1 Data Scarcity Bottleneck</a></li>
<li><a href="#sec5-2" class="toc-sublink">5.2 Dataset Partitioning & Dual Generators</a></li>
<li><a href="#sec5-3" class="toc-sublink">5.3 cmap Glyph Tofu Safeguard</a></li>
<li><a href="#sec5-4" class="toc-sublink">5.4 In-RAM Augmentations</a></li>
</ul>
</li>
<li class="toc-item">
<a href="#sec6" class="toc-link">6. Empirical Benchmarks</a>
<ul class="toc-sublist">
<li><a href="#sec6-1" class="toc-sublink">6.1 50-Test Benchmark (V23 vs V24)</a></li>
<li><a href="#sec6-2" class="toc-sublink">6.2 200-Page Stress-Test Suite</a></li>
<li><a href="#sec6-3" class="toc-sublink">6.3 Model Progression: V23 vs V24 vs V25 SOTA</a></li>
<li><a href="#sec6-4" class="toc-sublink">6.4 Multi-GPU Hardware Scaling</a></li>
</ul>
</li>
<li class="toc-item">
<a href="#sec7" class="toc-link">7. Production Serving</a>
<ul class="toc-sublist">
<li><a href="#sec7-1" class="toc-sublink">7.1 Microservice Architecture</a></li>
<li><a href="#sec7-2" class="toc-sublink">7.2 Concurrency & NumPy 2.x</a></li>
</ul>
</li>
<li class="toc-item">
<a href="#sec8" class="toc-link">8. Limitations & Open Challenges</a>
<ul class="toc-sublist">
<li><a href="#sec8-5" class="toc-sublink">8.5 Planned Post-V25 Iterations</a></li>
</ul>
</li>
<li class="toc-item"><a href="#sec9" class="toc-link">9. Citation & References</a></li>
</ul>
</aside>
<!-- Article Body -->
<article class="article-main">
<div class="paper-tag-strip">
<span class="tag-pill tag-pill-accent">Technical Report / Research Preprint</span>
<span class="tag-pill">Not Peer Reviewed</span>
<span class="tag-pill tag-pill-green">PaddleOCR PP-OCRv4</span>
<span class="tag-pill tag-pill-blue">Brahmic Paleography</span>
<span class="tag-pill">Unicode Eastern Cham (Akhar Thrah)</span>
</div>
<h1 class="paper-title">Cham-OCR: A Deep-Learning OCR Pipeline for Unicode Eastern Cham, toward Historical Manuscript and Epigraphic Recognition</h1>
<div class="author-box">
<div class="author-info">
<span class="author-name">Phuc H. Nguyen</span>
<span class="author-affiliation">Independent Researcher and Software Developer, Vietnam — Cham-OCR Project</span>
</div>
<div class="author-meta">
<span><strong>Published:</strong> September 2026</span>
<span><strong>Code & Weights:</strong> <a href="https://github.com/phucsd/Cham-OCR" target="_blank" rel="noopener noreferrer" style="color: var(--accent-primary); font-weight: 600;">GitHub phucsd/Cham-OCR</a></span>
<span><strong>Live Service:</strong> <a href="https://ocr.cham.asia" target="_blank" rel="noopener noreferrer" style="color: var(--accent-primary); font-weight: 600;">ocr.cham.asia</a></span>
<span><strong>Contact:</strong> <a href="mailto:phucsd@gmail.com" style="color: var(--accent-primary); font-weight: 600;">phucsd@gmail.com</a></span>
</div>
</div>
<!-- Abstract -->
<section id="abstract" class="abstract-box">
<div class="abstract-title">Abstract</div>
<p class="abstract-text">
Historical manuscripts of the Cham civilization—primarily preserved in Eastern Cham (<em>Akhar Thrah</em>) alongside Western Cham (<em>Cam Srak</em>) traditions—represent a fragile, millennia-old Brahmic epigraphic and literary heritage in Southeast Asia. While this project validates an end-to-end deep learning OCR pipeline specifically on Unicode Eastern Cham (<em>Akhar Thrah</em>), Western Cham (<em>Cam Srak</em>), real physical manuscripts, and epigraphic stone carvings represent active and ongoing future extensions. Automatic transcription via Optical Character Recognition (OCR) has historically faced steep barriers due to complex Brahmic abugida orthography, multi-directional dependent vowel diacritics, predominantly continuous script (<em>scriptio continua</em>) in classical manuscripts, physical carrier degradation, and tight interlinear diacritic collisions. In this technical report, we describe the design, implementation, and empirical evaluation of an end-to-end deep learning pipeline and diagnostic transcription workbench engineered for Cham documents: (1) a hybrid text detection and line segmentation engine combining a fine-tuned <strong>Mobile DBNet (PPLCNetV3)</strong> with an <strong>Indic Valley-Cut Heuristic Engine</strong> that preserves boundary strokes via mathematical set-difference masking (<code>difference_update</code>) and \(0.40 \times \text{median line height}\) vertical safety cushions; (2) a neural recognition backbone based on <strong>PP-OCRv4 SVTR-LCNet</strong> with expanded input dimensions \([3, 48, 480]\), joint Connectionist Temporal Classification (CTC) alignment, and post-OCR deterministic Unicode cluster canonicalization; and (3) a 150,000-textline synthetic dataset pipeline featuring on-the-fly motion blur, TrueType cmap tofu filtering, and hard-example adversarial mining for verse numerals (\(1\)–\(99\)) and confusing diacritic minimal pairs. Evaluations across a <strong>controlled 200-page synthetic document stress-test</strong> (903 textlines across 5 difficulty levels) and a <strong>50-test stratified failure-mode benchmark</strong> demonstrate robust textline detection with a <strong>100.0% line detection rate</strong> across standard, aged paper, and narrow-gap folios (Levels 1–3) and graceful degradation up to Level 4 (99.44% detection), before encountering an empirical breaking point at Level 5 (61.45% detection) where interlinear wave amplitude exceeds the interlinear gap. For textline recognition, our state-of-the-art model (<strong>V25</strong>) achieves <strong>90.96% Sequence Accuracy</strong> and <strong>99.28% Normalized Edit Distance (CER 0.72%)</strong> across 10,000 frozen validation samples after 40 epochs of multi-stage training on Dual Tesla T4x2 GPUs, significantly outperforming the V24 logical order baseline (88.94% accuracy) and V23 visual order legacy model. V25 successfully resolves Connectionist Temporal Classification (CTC) blank collapse on boundary punctuation (<span class="cham-script">꩝꩝</span> vs <span class="cham-script">꩝</span>), eliminates numeral-to-character misclassifications (<span class="cham-script">꩔</span> vs <span class="cham-script">ꨤ</span>, <span class="cham-script">꩕</span> vs <span class="cham-script">ꨅ</span>), and robustly disambiguates subtle diacritic minimal pairs (<span class="cham-script">ꨲ</span> vs <span class="cham-script">ꨶ</span>). Multi-GPU distributed scaling benchmarks across Kaggle Tesla T4x2, Cloud A100, L40S, and L4 characterize hardware efficiency for non-profit cultural preservation. Finally, we discuss critical limitations regarding synthetic-to-real domain gaps and provide a production reference deployed at <code>https://ocr.cham.asia</code>.
</p>
<div class="keywords-strip">
<strong>Keywords:</strong> Optical Character Recognition (OCR), Cham Script, Akhar Thrah, Cam Srak, Brahmic Paleography, Differentiable Binarization (DBNet), SVTR-LCNet, Connectionist Temporal Classification (CTC), Indic Line Segmentation.
</div>
</section>
<!-- SECTION 1 -->
<section id="sec1">
<h2>1. Introduction & Paleographic Foundations</h2>
<h3 id="sec1-1">1.1 Historical Script Typology & Document Endangered Status</h3>
<p>
The Cham language belongs to the Austronesian language family (Malayo-Polynesian branch) and was the primary medium of literature, state administration, and liturgy across the Champa polities along coastal central and southern Vietnam from the 2nd to early 19th centuries. The script family descends from Southern Brahmic Grantha/Pallava lineages, evolving into two distinct extant traditions:
</p>
<ul>
<li><strong>Eastern Cham (<em>Akhar Thrah</em>)</strong>: Inscribed and preserved predominantly in Ninh Thuận and Bình Thuận provinces (Vietnam). Historical manuscripts are executed on traditional bark-paper (<em>kertas</em>) or palm leaf with calligraphic bamboo pens.</li>
<li><strong>Western Cham (<em>Cam Srak</em>)</strong>: Preserved by Cham communities in the Mekong Delta (An Giang, Tây Ninh) and Cambodia. It features distinctive stylistic glyph contours, regional ligature forms, and localized phonetic loan signs.</li>
</ul>
<p>
Due to tropical climate conditions, ink bleed, microbiological decay, and physical fragmentation, remaining manuscript archives (such as the <em>Akayet Inra Patra</em>, <em>Ariya Po Pareng</em>, and sacred divinatory compendia) face significant preservation risks. Standard off-the-shelf OCR engines designed primarily for Latin or CJK scripts struggle significantly on historical Cham manuscripts due to fundamental typographical divergences.
</p>
<h3 id="sec1-2">1.2 Abugida Orthography & Multi-Directional Glyphs</h3>
<p>
Unlike alphabetic writing systems where letters follow a single horizontal axis, Cham is an <strong>abugida</strong> governed by complex spatial attachment rules:
</p>
<ol>
<li><strong>Independent Vowels (<code>U+AA00</code>–<code>U+AA05</code>)</strong>: Six standalone vowel glyphs represent syllable-initial vowel phonemes without a preceding consonant base:
<span class="cham-script">ꨀ</span> (U+AA00: <code>CHAM LETTER A</code>),
<span class="cham-script">ꨁ</span> (U+AA01: <code>CHAM LETTER I</code>),
<span class="cham-script">ꨂ</span> (U+AA02: <code>CHAM LETTER U</code>),
<span class="cham-script">ꨃ</span> (U+AA03: <code>CHAM LETTER E</code>),
<span class="cham-script">ꨄ</span> (U+AA04: <code>CHAM LETTER AI</code>),
<span class="cham-script">ꨅ</span> (U+AA05: <code>CHAM LETTER O</code>).
</li>
<li><strong>Consonantal Inventory (<code>U+AA06</code>–<code>U+AA28</code>)</strong>: Thirty-five consonant glyphs carry an inherent vowel (/a/ or /ɔ/), beginning with <span class="cham-script">ꨆ</span> (U+AA06: <code>CHAM LETTER KA</code>) through <span class="cham-script">ꨨ</span> (U+AA28: <code>CHAM LETTER HA</code>), encompassing unaspirated, aspirated, nasal, and liquid articulation series.</li>
<li><strong>Subjoined Medials (<code>U+AA33</code>–<code>U+AA36</code>)</strong>: Four medial consonants attach underneath or encircle the consonant base:
Medial Ya (<span class="cham-script">ꨳ</span> U+AA33: <code>CHAM CONSONANT SIGN YA</code>),
Medial Ra (<span class="cham-script">ꨴ</span> U+AA34: <code>CHAM CONSONANT SIGN RA</code>),
Medial La (<span class="cham-script">ꨵ</span> U+AA35: <code>CHAM CONSONANT SIGN LA</code>), and
Medial Wa (<span class="cham-script">ꨶ</span> U+AA36: <code>CHAM CONSONANT SIGN WA</code>).
</li>
<li><strong>Multi-Directional Dependent Vowels (<code>U+AA29</code>–<code>U+AA32</code>)</strong>: Dependent vowel signs attach across all four spatial quadrants around the base:
<ul>
<li><em>Above base</em>: Vowel Sign I (<span class="cham-script">ꨪ</span> U+AA2A), II (<span class="cham-script">ꨫ</span> U+AA2B), EI (<span class="cham-script">ꨬ</span> U+AA2C).</li>
<li><em>Below base</em>: Vowel Sign U (<span class="cham-script">ꨭ</span> U+AA2D), UE (<span class="cham-script">ꨲ</span> U+AA32).</li>
<li><em>Post-base / Dependent</em>: Vowel Sign AA (<span class="cham-script">ꨩ</span> U+AA29), OE (<span class="cham-script">ꨮ</span> U+AA2E), AU (<span class="cham-script">ꨱ</span> U+AA31).</li>
<li><em>Pre-base (Left / Visual Ahead)</em>: Vowel Sign O (<span class="cham-script">ꨯ</span> U+AA2F), AI (<span class="cham-script">ꨰ</span> U+AA30).</li>
</ul>
</li>
<li><strong>Final Consonants & Signs ([<code>U+AA40</code>–<code>U+AA4D</code>, <code>U+AA25</code>])</strong>: Explicit syllable codas including separate base glyphs with extended terminal strokes (<code>U+AA40</code>–<code>U+AA4B</code>), combining coda marks (Final Ng <span class="cham-script">ꩃ</span> U+AA43, Final M / Anusvara <span class="cham-script">ꩌ</span> U+AA4C, Final H / Visarga <span class="cham-script">ꩍ</span> U+AA4D), and the base consonant <span class="cham-script">ꨥ</span> (<code>U+AA25</code> CHAM LETTER VA) which, per Unicode Standard Chapter 16 Table 16-16, functions without alteration in both initial and syllable-final positions (e.g., in syllables such as <span class="cham-script">ꨀꨍꨯꨱꨥ</span> and <span class="cham-script">ꨝꨗꨴꨭꨥ</span>).</li>
<li><strong>Script Spacing Conventions</strong>: Many historical Cham manuscripts exhibit limited or inconsistent inter-word spacing, while modern printed Cham commonly uses spaces. Classical literary texts and verse compendia frequently exhibit continuous running script (<em>scriptio continua</em>), though section divisions, poetic hemistichs, or phrasing pauses are demarcated by punctuation marks (e.g., Danda <span class="cham-script">꩝</span>, Double Danda <span class="cham-script">꩞</span>, Triple Danda <span class="cham-script">꩟</span>) and occasional whitespace gaps.</li>
<li><strong>Punctuation & Verse Markers</strong>: Traditional discourse is demarcated via Danda (<span class="cham-script">꩝</span> U+AA5D), Double Danda (<span class="cham-script">꩞</span> U+AA5E, also serving as section/stanza mark), and Triple Danda (<span class="cham-script">꩟</span> U+AA5F). Poetic stanzas frequently open with Cham numerals followed by a section mark (e.g., <span class="cham-script">꩑꩞</span> for Verse 1, <span class="cham-script">꩒꩞</span> for Verse 2).</li>
</ol>
<h3 id="sec1-3">1.3 The Conflict: Logical Storage Order vs Visual Rendering Order</h3>
<p>
The Unicode Standard mandates <strong>Logical Order</strong> encoding for all Brahmic scripts, meaning the character sequence reflects phonetic utterance:
</p>
<div class="math-block">
$$\text{Canonical Cluster} = \text{Base Consonant} + [\text{Medials}] + [\text{Pre-Vowels}] + [\text{Dependent Vowels}] + [\text{Finals / Signs}]$$
</div>
<p>
However, pre-vowels like <span class="cham-script">ꨯ</span> (Vowel Sign O, U+AA2F) and <span class="cham-script">ꨰ</span> (Vowel Sign AI, U+AA30) are rendered <em>visually to the left</em> of the base consonant. When an optical sequence decoder scans an image crop left-to-right, its visual perception encounters the pre-vowel ahead of the consonant glyph (e.g. perceiving <span class="cham-script">ꨯ</span> before base <span class="cham-script">ꨕ</span>). Standard CTC models therefore frequently output the illegal sequence <span class="cham-script">ꨯꨕ</span> instead of the canonical Unicode cluster <span class="cham-script">ꨕꨯ</span>.
</p>
<div class="academic-callout callout-warning">
<div class="callout-title">⚠️ Paleographic Pitfall: Visual Parsing Corruption</div>
Attempting to resolve visual ordering with naive string replacements or visual-to-unicode cluster splitters damages valid multi-diacritic compounds containing subjoined medials (such as <span class="cham-script">ꨙꨳꨯꨮ</span>). Our system strictly mandates a deterministic post-OCR <strong>Logical Order Normalizer</strong> (Section 4.3) that re-aligns local visual permutations into standardized Unicode sequences without corrupting consonant-medial clusters.
</div>
</section>
<!-- SECTION 2: RELATED WORK -->
<section id="sec2">
<h2>2. Related Work on Cham Document Analysis & Epigraphy</h2>
<h3 id="sec2-1">2.1 Early Epigraphic & Glyph Recognition</h3>
<p>
Automated processing of Cham script originated in the analysis of stone inscriptions from ancient Champa sanctuaries (Mỹ Sơn, Po Nagar, Đồng Dương). Nguyen, Schweyer, Le, Tran, & Vu (2019a) presented preliminary results on recognizing ancient Cham glyphs extracted from stone rubbings and photographs using Histogram of Oriented Gradients (HOG), Non-Parametric Weighted (NPW) features, and CNN-derived feature representations coupled with $k$-NN and linear SVM classifiers at the IEEE MAPR 2019 conference. Due to severe stone weathering, erosion, and irregular surface texture, isolated glyph classification across noisy background substrates was evaluated.
</p>
<p>
To mitigate extreme sample scarcity in epigraphic corpora, Nguyen, Schweyer, Le, Tran, & Vu (2019b) evaluated data augmentation strategies and transfer learning from scripts of similar or the same language family to ancient stone inscription glyphs at the ICIAP 2019 Workshops (PatReCH), demonstrating quantitative gains on script-to-inscription transfer across weathered stone carvings.
</p>
<h3 id="sec2-2">2.2 Deep Learning for Cham Epigraphy & Transliteration</h3>
<p>
In his doctoral dissertation, Nguyen (2023) developed comprehensive document image processing and understanding methodologies dedicated to ancient Cham documents. His work investigated stone inscription image restoration, automatic glyph segmentation under rough stone substrate conditions, and deep convolutional representations for historical Brahmic scripts.
</p>
<p>
Complementing glyph recognition, Nguyen, Burie, Le, & Schweyer (2022) proposed an effective multi-stage method for text line segmentation in historical document images presented at the 26th International Conference on Pattern Recognition (ICPR 2022), addressing line curvature and touching diacritics in historical archives. Expanding to linguistic representation, Nguyen, Burie, Le, & Schweyer (2023) introduced a two-step sequence transformer model for Cham-to-Latin script transliteration presented at the HIP@ICDAR 2023 workshop. Their system demonstrated the efficacy of sequence-to-sequence transformers in mapping recognized Cham text sequences to standardized Latin phonological romanizations (<em>Cam-Latin</em>). Most recently, Nguyen, Burie, Le, & Schweyer (2025) published a comprehensive study entitled <em>Text line segmentation approach combining deep learning model and traditional image processing techniques - application to transliteration of Cham manuscripts</em> in <em>Multimedia Tools and Applications</em>, evaluating text-line segmentation and automatic transliteration across 627 manuscript images and approximately 8,300 textlines (focusing on line segmentation and transliteration rather than end-to-end character OCR benchmarking).
</p>
<h3 id="sec2-3">2.3 Manuscript Preservation & Digitization Initiatives</h3>
<p>
Beyond epigraphy on stone, historical Cham manuscripts on paper (<em>kertas</em>) and palm-leaf have been documented through preservation missions. The Historic Cham Manuscripts of Vietnam collection in the Southeast Asia Digital Library (SEADL) contains 977 digitized manuscripts comprising more than 57,800 page scans. In addition, field surveys estimate approximately 3,000 physical manuscripts preserved within Cham community families and religious dignitaries across Vietnam and Cambodia. The French National Research Agency funded the CHAMDOC project (<em>Cham Documentation</em>, ANR-19-CE27-0018, 2019–2024) to support preservation, cataloging, and linguistic documentation.
</p>
<p>
However, despite the preservation of tens of thousands of scanned folios, almost none of these archives possess open, machine-readable character-level or line-level bounding box ground-truth annotations suitable for training supervised deep learning models. The bottleneck for Cham OCR is therefore not an absolute absence of physical manuscripts, but rather the severe scarcity of annotated, verified training ground truth.
</p>
<h3 id="sec2-4">2.4 Positioning of the Present System</h3>
<p>
Prior research has focused predominantly on either <strong>isolated glyph recognition on stone inscriptions</strong>, <strong>downstream transliteration of cleaned text</strong>, or academic research prototypes evaluated on private manuscript subsets. In contrast, the system presented in this report addresses the complete <strong>end-to-end continuous document pipeline</strong> for historical manuscripts:
</p>
<ul>
<li>Full-page line detection and segmentation robust to undulating baselines and tight interlinear diacritics.</li>
<li>End-to-end continuous textline recognition without manual character pre-segmentation.</li>
<li>Deterministic canonical reordering adhering strictly to the official Unicode standard.</li>
<li>Production-grade serving designed for cultural heritage researchers and native community access.</li>
</ul>
</section>
<!-- Section 3 -->
<section id="sec3">
<h2>3. Text Detection & Indic Line Segmentation (Cham-DBNet)</h2>
<h3 id="sec3-1">3.1 The Interlinear Diacritic Collision Problem</h3>
<p>
In digital modeling and synthetic text rendering at nominal glyph heights (32–36px), interlinear clearance often narrows to 3–12 pixels between adjacent lines (with ascender excursions extending up to \(+18\text{px}\) and descender excursions down to \(-22\text{px}\) in normalized coordinate space). Because upper dependent vowels reach high above the core text band and subjoined medials plunge below, the ascenders of line \(n+1\) routinely touch or intertwine with the descenders of line \(n\) under dense formatting. Standard off-the-shelf text detection models (such as baseline DBNet trained on natural scene Latin or CJK text) frequently encounter two severe failure modes when applied to dense historical manuscripts:
</p>
<ul>
<li>Merging adjacent lines into single monolithic boxes.</li>
<li>Slicing off upper/lower diacritics, leading to fatal recognition errors.</li>
</ul>
<h3 id="sec3-2">3.2 Specialized Cham-DBNet: Architecture & Training</h3>
<p>
We trained a specialized text detector based on <strong>PP-OCRv4 Mobile DBNet</strong> with a lightweight <strong>PPLCNetV3</strong> backbone on an NVIDIA L4 GPU (24GB VRAM) in Lightning AI:
</p>
<div class="table-responsive">
<table>
<thead>
<tr>
<th>Hyperparameter / Component</th>
<th>Specification</th>
<th>Engineering Rationale</th>
</tr>
</thead>
<tbody>
<tr>
<td>Backbone Network</td>
<td><code>PPLCNetV3</code> (scale 0.75)</td>
<td>Depthwise separable conv + SE channel attention; ultra-fast inference on CPU/Edge.</td>
</tr>
<tr>
<td>Loss Function</td>
<td>Probability Map Loss (\(L_s\)) + Binary Loss (\(L_b\)) + Threshold Loss (\(L_{th}\))</td>
<td>Differentiable binarization with adaptive threshold learning (Dice Loss).</td>
</tr>
<tr>
<td>Unclip Ratio (\(\alpha\))</td>
<td><code>1.8</code></td>
<td>Higher than Latin default (1.5) to capture high ascenders and long subjoined medials.</td>
</tr>
<tr>
<td>Binarization Threshold</td>
<td><code>0.25</code></td>
<td>Tuned for faint, faded ink on historical paper bark.</td>
</tr>
<tr>
<td>Polygon Box Threshold</td>
<td><code>0.50</code></td>
<td>Filters out stray ink specks and background parchment artifacts.</td>
</tr>
<tr>
<td>Max Candidates</td>
<td><code>1500</code></td>
<td>Accommodates densely packed multi-line folios without candidate truncation.</td>
</tr>
<tr>
<td>Training Dataset</td>
<td>2,000 document pages</td>
<td>50% synthesized with narrow interline spacing (3–10px) to penalize box merging (as configured in <code>launch_h100_full_pipeline.py</code>).</td>
</tr>
<tr>
<td>Optimizer & Schedule</td>
<td>Adam (Cosine decay, initial lr 0.001, 5 warmup epochs, L2 factor 5e-5)</td>
<td>Stable gradient convergence without aggressive parameter divergence.</td>
</tr>
</tbody>
</table>
</div>
<p>
<strong>Empirical Line Detection Performance</strong>: Evaluated within the complete document processing pipeline on our 200-page synthetic stress-test suite (<code>evaluation_report.json</code>), Cham-DBNet achieves high line extraction fidelity across difficulty levels:
</p>
<div class="table-responsive">
<table>
<thead>
<tr>
<th>Stress Tier / Evaluation Condition</th>
<th class="num">GT Lines</th>
<th class="num">Detected</th>
<th class="num">Line Detection Rate</th>
<th>Empirical Observation</th>
</tr>
</thead>
<tbody>
<tr>
<td>Level 1: Standard (Gap 25–35px)</td>
<td class="num">201</td>
<td class="num">201</td>
<td class="num"><span class="metric-badge metric-good">100.0%</span></td>
<td>Flawless detection across clean rectilinear textlines.</td>
</tr>
<tr>
<td>Level 2: Aged Paper (Noise, Skew ≤ 2°)</td>
<td class="num">189</td>
<td class="num">189</td>
<td class="num"><span class="metric-badge metric-good">100.0%</span></td>
<td>Resilient to aged background texture and paper fiber noise.</td>
</tr>
<tr>
<td>Level 3: Narrow Gap (Gap 8–14px)</td>
<td class="num">170</td>
<td class="num">170</td>
<td class="num"><span class="metric-badge metric-good">100.0%</span></td>
<td>Zero line merging on tightly spaced manuscript lines.</td>
</tr>
<tr>
<td>Level 4: Wavy Sinusoid (Warping 3–5.5px)</td>
<td class="num">177</td>
<td class="num">176</td>
<td class="num"><span class="metric-badge metric-good">99.44%</span></td>
<td>Graceful degradation under undulating non-linear baselines.</td>
</tr>
<tr>
<td>Level 5: Extreme Overlap (Amplitude > Gap)</td>
<td class="num">166</td>
<td class="num">102</td>
<td class="num"><span class="metric-badge metric-warn">61.45%</span></td>
<td>Empirical breaking point where intersecting strokes cause line merging.</td>
</tr>
<tr>
<td><strong>Overall Suite Summary</strong></td>
<td class="num"><strong>903</strong></td>
<td class="num"><strong>838</strong></td>
<td class="num"><strong><span class="metric-badge metric-good">92.80%</span></strong></td>
<td><strong>Average end-to-end CPU latency: 1.33s / page</strong></td>
</tr>
</tbody>
</table>
</div>
<h3 id="sec3-3">3.3 Hybrid Indic Valley-Cut Heuristic & Boundary Masking Safeguards</h3>
<p>
For CPU execution in low-resource environments (or when DBNet probability maps exhibit ambiguity), we designed an <strong>Indic Valley-Cut Heuristic Engine</strong> operating as a secondary defense layer:
</p>
<ol>
<li><strong>Connected Component Analysis (CCA)</strong>: Extracts all connected ink components on the binarized page.</li>
<li><strong>Core Text Band Determination</strong>: Identifies the median baseline and core bounding box of each text line using horizontal projection histograms.</li>
<li><strong>Boundary Stroke Masking via <code>difference_update</code></strong>: When creating an erasing mask to obscure adjacent interfering lines, components belonging to the current line are strictly excluded using mathematical set difference (\(S_{\text{mask}} = S_{\text{adjacent}} \setminus S_{\text{current}}\)). This prevents amputating touching diacritics.</li>
<li><strong>Vertical Padding Safeguards</strong>: The bounding box is padded by exactly \(0.40 \times \text{median line height}\) upwards and downwards, bounded at a minimum distance of \(2\text{px}\) from neighboring core bands.</li>
<li><strong>Legacy-First / No-Change Gate</strong>: For clean documents where interline gap \(\ge 0.35 \times \text{height}\) and initial confidence \(\ge 0.75\), multi-crop splitting is bypassed entirely to guarantee zero regression on pristine folios.</li>
</ol>
</section>
<!-- SECTION 4 -->
<section id="sec4">
<h2>4. Neural Text Recognition Architecture (Cham-SVTR)</h2>
<h3 id="sec4-1">4.1 PP-OCRv4 SVTR-LCNet Backbone</h3>
<p>
Text recognition is formulated as sequence labeling on textline image crops. We adopted the <strong>SVTR-LCNet</strong> architecture from PP-OCRv4, customized with specialized input and output heads:
</p>
<ul>
<li><strong>Input Tensor Dimensions</strong>: Scaled to \([3, 48, 480]\) (channels \(\times\) height \(\times\) width). Widening the receptive field to 480 width prevents spatial squashing of Cham verses containing stacked diacritics.</li>
<li><strong>SVTR Blocks (Single Visual Model for Text Recognition)</strong>: Integrates local visual feature extractors with patch-wise self-attention mechanisms to capture long-range contextual relationships across unspaced text.</li>
<li><strong>Squeeze-and-Excitation (SE) Channel Attention</strong>: Adaptively recalibrates channel-wise feature responses, emphasizing subtle diacritic strokes over background paper grain.</li>
</ul>
<h3 id="sec4-2">4.2 Joint CTC & Sequence Attention Loss</h3>
<p>
Training employs joint supervision combining Connectionist Temporal Classification (CTC) with an Attention-based sequence decoder:
</p>
<div class="math-block">
$$\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{CTC}} + \lambda \mathcal{L}_{\text{attention}}$$
</div>
<p>
Where \(\mathcal{L}_{\text{CTC}}\) maximizes the log-likelihood of valid alignments over the transcription sequence \(\mathbf{y}\):
</p>
<div class="math-block">
$$\mathcal{L}_{\text{CTC}} = -\ln P(\mathbf{y} \mid \mathbf{x}) = -\ln \sum_{\pi \in \mathcal{B}^{-1}(\mathbf{y})} P(\pi \mid \mathbf{x})$$
</div>
<p>
During runtime inference, greedy CTC decoding is executed to provide lightweight, low-latency single-thread CPU execution without costly beam search.
</p>
<h3 id="sec4-3">4.3 Post-OCR Deterministic Canonical Normalization</h3>
<p>
To resolve the visual-versus-logical ordering paradox described in Section 1.3, we developed <code>normalize_unicode</code>, an algorithmic state machine built upon <code>parse_unicode_clusters</code>. It decomposes the raw CTC output string into individual phonological syllables and re-sorts tokens into canonical Brahmic sequence:
</p>
<div class="code-box">
<div class="code-header">
<span>Python Algorithm: normalize_unicode</span>
<button class="copy-btn" onclick="navigator.clipboard.writeText(document.getElementById('norm-code').innerText)">Copy</button>
</div>
<pre id="norm-code"><code>def normalize_unicode(text: str) -> str:
# Enforces Canonical Brahmic Logical Order on CTC decoded Cham text:
# Base Consonant -> Medial RA/LA -> Medial YA/WA -> Pre-Vowels (O/AI) ->
# Other Dependent Vowels -> Vowel Lengthener AA -> Final Consonants / Signs (incl. Final VA)
# Example: Corrects visual sequence 'ꨙꨯꨳꨮ' into canonical 'ꨙꨳꨯꨮ'.
clusters = parse_unicode_clusters(text)
uni_clusters = [visual_to_unicode_cluster(c) for c in clusters]
return "".join("".join(c) for c in uni_clusters)</code></pre>
</div>
</section>
<!-- SECTION 5 -->
<section id="sec5">
<h2>5. Synthetic Dataset Generation & Hard-Example Mining</h2>
<h3 id="sec5-1">5.1 The Data Scarcity Bottleneck</h3>
<p>
Deep neural OCR networks require hundreds of thousands of diverse labeled textlines. However, surviving physical Cham manuscripts lack open character-level bounding box ground truth. To overcome this limitation, we engineered an algorithmic synthesizer (<code>scripts/generate_data.py</code>) capable of rendering 150,000 diverse textline crops from authentic literary corpora.
</p>
<h3 id="sec5-2">5.2 Calibrated Synthetic Dataset Partitioning & Dual-Generator Architecture</h3>
<p>
To support both baseline reproducibility and ongoing experimental modeling, the repository provides two distinct synthetic dataset generation pipelines:
</p>
<h4 style="margin-top: 1.25rem; font-size: 1.05rem; color: var(--text-primary);">Pipeline A: Baseline General Generator (<code>scripts/generate_data.py</code>)</h4>
<p>
The baseline generator produces a 150,000 textline corpus partitioned into <strong>15% evaluation/locked splits</strong> (22,500 lines: <code>val_clean_short</code> 3%, <code>val_clean_long</code> 3%, <code>val_noisy_short</code> 3%, <code>val_noisy_long</code> 3%, <code>locked_test</code> 3%) and an <strong>85% training split</strong> (127,500 lines) structured across 6 calibrated sub-categories:
</p>
<div class="table-responsive">
<table>
<thead>
<tr>
<th>Training Sub-Category</th>
<th class="num">Proportion</th>
<th class="num">Line Count</th>
<th>Composition & Generation Protocol</th>
<th>Target Failure Mode Addressed</th>
</tr>
</thead>
<tbody>
<tr>
<td><code>train_clean_short</code></td>
<td class="num">25%</td>
<td class="num">31,875</td>
<td>Short clean literary phrases (1–4 tokens).</td>
<td>Fast convergence on high-frequency base consonants and common syllables.</td>
</tr>
<tr>
<td><code>train_clean_medium</code></td>
<td class="num">20%</td>
<td class="num">25,500</td>
<td>Medium-length lines (5–8 tokens) from classical texts (<em>Akayet Inra Patra</em>, <em>Ariya Po Pareng</em>).</td>
<td>Teaches authentic Cham phonotactic and n-gram transitions.</td>
</tr>
<tr>
<td><code>train_clean_long</code></td>
<td class="num">15%</td>
<td class="num">19,125</td>
<td>Long full verse hemistichs and prose sentences (9+ tokens).</td>
<td>Robustness to full-line receptive fields and CTC sequence alignment without truncation.</td>
</tr>
<tr>
<td><code>train_noisy_short</code></td>
<td class="num">15%</td>
<td class="num">19,125</td>
<td>Short lines rendered with Point Spread Function (PSF) blur, noise, and lighting gradients.</td>
<td>Invariance to archival handling artifacts, tremor, and fading ink.</td>
</tr>
<tr>
<td><code>train_noisy_long</code></td>
<td class="num">15%</td>
<td class="num">19,125</td>
<td>Long verse lines with severe photometric degradation and spatial elastic distortion.</td>
<td>Prevents alignment collapse on degraded multi-word lines.</td>
</tr>
<tr>
<td><code>train_hard_examples</code></td>
<td class="num">10%</td>
<td class="num">12,750</td>
<td>Adversarial minimal pairs (<span class="cham-script">ꨲ</span> vs <span class="cham-script">ꨶ</span>, <span class="cham-script">꩝꩝</span> vs <span class="cham-script">꩝</span>), verse numerals (1–99) with section marks (<span class="cham-script">꩑꩞</span>), and stacked subjoined clusters.</td>
<td>Eliminates CTC misclassification of numerals as letters (<span class="cham-script">꩔</span> vs <span class="cham-script">ꨤ</span>, <span class="cham-script">꩕</span> vs <span class="cham-script">ꨅ</span>) and diacritic collapsing.</td>
</tr>
</tbody>
</table>
</div>
<h4 style="margin-top: 1.5rem; font-size: 1.05rem; color: var(--text-primary);">Pipeline B: Experimental V25 Generator (<code>scripts/generate_data_v25.py</code>)</h4>
<p>
For the V25 experimental iteration, a dedicated multi-threaded generator (<code>generate_data_v25.py</code>) synthesizes 150,000 textline images (140,000 train + 10,000 validation) directly synchronized with <code>configs/v25_dataset_manifest.json</code> across 5 core pillars:
</p>
<ol style="margin-bottom: 1.5rem; line-height: 1.7; color: var(--text-secondary);">
<li><strong>Canonical Cham Literature (43.3% / 65,000 lines)</strong>: Classical literary vocabulary (<em>Po Klong Garai</em>, 57-stanza verse).</li>
<li><strong>Anti-Motion Blur Base (20.0% / 30,000 lines)</strong>: Clean base images paired with dynamic in-RAM directional motion blur (\(7\times 7\) to \(13\times 13\)) and defocus.</li>
<li><strong>Inline Bilingual Code-Switching (16.7% / 25,000 lines)</strong>: Intra-line mixed phrases with Vietnamese diacritics and Latin annotations.</li>
<li><strong>Stanza Numerals & Boundary Punctuation (12.0% / 18,000 lines)</strong>: Verse numbers \(1\)–\(99\) (<span class="cham-script">꩑꩞</span>..<span class="cham-script">꩙꩙꩞</span>) and boundary marks (<span class="cham-script">꩞</span>, <code>:</code>, <code>–</code>).</li>
<li><strong>Adversarial Minimal Pairs (8.0% / 12,000 lines)</strong>: Micro-stroke pairs (<span class="cham-script">ꨲ</span> U+AA32 Vowel Sign UE vs <span class="cham-script">ꨶ</span> U+AA36 Medial WA, variable Double Danda <span class="cham-script">꩝꩝</span> spacing 2–8px, 3-tier diacritic stacks).</li>
</ol>
<div class="academic-callout callout-info">
<div class="callout-title">✅ Completed Multi-Stage Training (V25)</div>
Model Version 25 has completed its full 40-epoch multi-stage distributed GPU training on Kaggle Dual Tesla T4x2 accelerators (Stage 1: Epochs 1–20 warmup and baseline adaptation; Stage 2: Epochs 21–30 hard-example fine-tuning; Stage 3: Epochs 31–40 final convergence with cosine learning rate decay, completing 87,480 cumulative training steps). The final checkpoint (<code>best_accuracy.pdparams</code>, 58.84 MB) achieved <strong>90.96% Sequence Accuracy</strong> and <strong>99.28% Normalized Edit Distance (CER 0.72%)</strong> on the 10,000 frozen validation suite.
</div>
<h3 id="sec5-3">5.3 Digital Typefaces & TrueType cmap Tofu Glyph Safeguards</h3>
<p>
When rendering synthetic text, digital Cham typefaces must be verified for glyph coverage. The primary bundled and reproducible typeface distributed directly within this repository is <strong>Noto Sans Cham</strong> (<code>Regular</code>, <code>Bold</code>, <code>Black</code>; SIL Open Font License 1.1, located in <code>ocr-training/data/fonts/</code>). External reference typefaces evaluated during research—including <em>Cham Roman</em>, <em>EFEO Cham</em>, and community fonts—are not redistributed in this repository and must be acquired independently if researchers wish to reproduce specific historical font variations.
</p>
<p>
To prevent missing Unicode codepoints from producing blank rectangles ("tofu" glyphs), our generator incorporates pre-render verification via <code>fontTools.ttLib</code> to extract each font's <code>cmap</code> table. Candidate strings are inspected character-by-character; any string containing unmapped glyphs for the active font is rejected prior to rasterization.
</p>
<h3 id="sec5-4">5.4 On-the-Fly Dynamic Motion Blur & Noise Synthesis in RAM</h3>
<p>
To prevent disk I/O bottlenecks during generation, augmentations are synthesized dynamically in memory:
</p>
<ul>
<li><strong>Point Spread Function (PSF) Motion Blur</strong>: Simulates hand tremors during mobile photography in archival field trips (kernel size 3–9px, angle \([0, 180^\circ]\)).</li>
<li><strong>Gaussian & Salt-and-Pepper Noise</strong>: Simulates coarse paper fibers and dust particles.</li>
<li><strong>Non-Linear Illumination Gradients</strong>: Models uneven natural sunlight across warped manuscript pages.</li>
</ul>
</section>
<!-- SECTION 6 -->
<section id="sec6">
<h2>6. Empirical Benchmarks & Quantitative Evaluations</h2>
<h3 id="sec6-1">6.1 Controlled 50-Test Stratified Benchmark (Model V23 vs V24)</h3>
<p>
We evaluated model evolution across 50 rigorous stratified test cases covering 10 distinct failure modes, logged in <code>benchmark_50_tests_results.json</code> and <code>benchmark_50_tests_results_v24.json</code>:
</p>
<div class="table-responsive">
<table>
<thead>
<tr>
<th>Evaluation Category</th>
<th class="num">Count</th>
<th class="num">V23 CER (%)</th>
<th class="num">V24 CER (%)</th>
<th class="num">CER Gain</th>
<th class="num">V23 Pass (%)</th>
<th class="num">V24 Pass (%)</th>
<th class="num">V24 Conf</th>
</tr>
</thead>
<tbody>
<tr>
<td>Cat 1: Clean Baseline</td>
<td class="num">5</td>
<td class="num">14.84%</td>
<td class="num"><strong>6.20%</strong></td>
<td class="num"><span class="metric-badge metric-good">-8.64%</span></td>
<td class="num">20.0%</td>
<td class="num"><strong>80.0%</strong></td>
<td class="num">0.958</td>
</tr>
<tr>
<td>Cat 2: Single-Digit Stanzas</td>
<td class="num">5</td>
<td class="num">18.42%</td>
<td class="num"><strong>11.71%</strong></td>
<td class="num"><span class="metric-badge metric-good">-6.71%</span></td>
<td class="num">0.0%</td>
<td class="num"><strong>40.0%</strong></td>
<td class="num">0.944</td>
</tr>
<tr>
<td>Cat 3: Two-Digit Stanzas</td>
<td class="num">5</td>
<td class="num">17.59%</td>
<td class="num"><strong>7.65%</strong></td>
<td class="num"><span class="metric-badge metric-good">-9.94%</span></td>
<td class="num">20.0%</td>
<td class="num"><strong>60.0%</strong></td>
<td class="num">0.956</td>
</tr>
<tr>
<td>Cat 4: Low Height / Low-Res</td>
<td class="num">5</td>
<td class="num">27.65%</td>
<td class="num"><strong>17.96%</strong></td>
<td class="num"><span class="metric-badge metric-good">-9.69%</span></td>
<td class="num">0.0%</td>
<td class="num"><strong>20.0%</strong></td>
<td class="num">0.848</td>
</tr>
<tr>
<td>Cat 5: Blur Degradation</td>
<td class="num">5</td>
<td class="num">41.78%</td>
<td class="num"><strong>35.56%</strong></td>
<td class="num"><span class="metric-badge metric-good">-6.22%</span></td>
<td class="num">0.0%</td>
<td class="num">0.0%</td>
<td class="num">0.839</td>
</tr>
<tr>
<td>Cat 6: Noise & Grain</td>
<td class="num">5</td>
<td class="num">50.59%</td>
<td class="num"><strong>8.82%</strong></td>
<td class="num"><span class="metric-badge metric-good">-41.77%</span></td>
<td class="num">0.0%</td>
<td class="num"><strong>80.0%</strong></td>
<td class="num">0.944</td>
</tr>
<tr>
<td>Cat 7: Paper Texture & Contrast</td>
<td class="num">5</td>
<td class="num">33.33%</td>
<td class="num"><strong>21.33%</strong></td>
<td class="num"><span class="metric-badge metric-good">-12.00%</span></td>
<td class="num">0.0%</td>
<td class="num">0.0%</td>
<td class="num">0.921</td>
</tr>
<tr>
<td>Cat 8: Tilt & Perspective</td>
<td class="num">5</td>
<td class="num">22.14%</td>
<td class="num"><strong>0.71%</strong></td>
<td class="num"><span class="metric-badge metric-good">-21.43%</span></td>
<td class="num">0.0%</td>
<td class="num"><strong>100.0%</strong></td>
<td class="num">0.982</td>
</tr>
<tr>
<td>Cat 9: Stroke Degradation</td>
<td class="num">5</td>
<td class="num">26.67%</td>
<td class="num"><strong>6.67%</strong></td>
<td class="num"><span class="metric-badge metric-good">-20.00%</span></td>
<td class="num">0.0%</td>
<td class="num"><strong>80.0%</strong></td>
<td class="num">0.928</td>
</tr>
<tr>
<td>Cat 10: Diacritics & Punctuation</td>
<td class="num">5</td>
<td class="num">28.72%</td>
<td class="num"><strong>22.19%</strong></td>
<td class="num"><span class="metric-badge metric-good">-6.53%</span></td>
<td class="num">0.0%</td>
<td class="num"><strong>40.0%</strong></td>
<td class="num">0.913</td>
</tr>
<tr style="font-weight: 700; background: var(--bg-subtle);">
<td>Overall Benchmark Average</td>
<td class="num">50</td>
<td class="num">28.17%</td>
<td class="num"><strong>13.88%</strong></td>
<td class="num"><span class="metric-badge metric-good">-14.29%</span></td>
<td class="num">4.0%</td>
<td class="num"><strong>50.0%</strong></td>
<td class="num">0.923</td>
</tr>
</tbody>
</table>
</div>
<h3 id="sec6-2">6.2 Controlled 200-Page Synthetic Document Stress-Test (903 Textlines)</h3>
<p>
To establish the operational limits of the complete end-to-end pipeline (Cham-DBNet + Cham-SVTR v24), we evaluated a <strong>controlled synthetic stress-test suite</strong> of 200 generated page images comprising 903 ground-truth textlines. Generated by <code>ocr-benchmark/scripts/generate_benchmark_200.py</code>, this suite renders authentic literary passages under calibrated geometric and photometric stress across 5 difficulty tiers:
</p>
<div class="table-responsive">
<table>
<thead>
<tr>
<th>Difficulty Tier</th>
<th>Stress & Degradation Profile</th>
<th class="num">Samples</th>
<th class="num">GT Lines</th>
<th class="num">Detection Rate</th>
<th class="num">Mean CER</th>
<th class="num">Mean WER</th>
<th class="num">CPU Latency</th>
</tr>
</thead>
<tbody>
<tr>
<td>Level 1: Standard</td>
<td>Line gap 25–35px, rectilinear alignment, white background.</td>
<td class="num">40</td>
<td class="num">201</td>
<td class="num"><span class="metric-badge metric-good">100.0%</span></td>
<td class="num">14.43%</td>
<td class="num">45.50%</td>
<td class="num">1.68s</td>
</tr>
<tr>
<td>Level 2: Aged Paper</td>
<td>Yellowed paper texture, mild dust grain, slight skew (\(\le 2^\circ\)).</td>
<td class="num">40</td>
<td class="num">189</td>
<td class="num"><span class="metric-badge metric-good">100.0%</span></td>
<td class="num">14.44%</td>
<td class="num">45.86%</td>
<td class="num">1.50s</td>
</tr>
<tr>
<td>Level 3: Narrow Gap</td>
<td>Tight interline gap 8–14px, micro-waviness (1–2px).</td>
<td class="num">40</td>
<td class="num">170</td>
<td class="num"><span class="metric-badge metric-good">100.0%</span></td>
<td class="num">12.77%</td>
<td class="num">43.52%</td>
<td class="num">1.26s</td>
</tr>
<tr>
<td>Level 4: Wavy Sinusoid</td>
<td>Sine wave warping (amp 3–5.5px), perspective tilt (0.03–0.05), gap 5–10px.</td>
<td class="num">40</td>
<td class="num">177</td>
<td class="num"><span class="metric-badge metric-good">99.44%</span></td>
<td class="num">14.53%</td>
<td class="num">45.64%</td>
<td class="num">1.24s</td>
</tr>
<tr>
<td>Level 5: Extreme Overlap</td>
<td>Gap 2–6px < wave amplitude 5.5–8.0px. Intersecting boundary strokes.</td>
<td class="num">40</td>
<td class="num">166</td>
<td class="num"><span class="metric-badge metric-bad">61.45%</span></td>
<td class="num">60.02%</td>
<td class="num">77.63%</td>
<td class="num">0.99s</td>
</tr>
<tr style="font-weight: 700; background: var(--bg-subtle);">
<td>Overall Suite Summary</td>
<td>200 synthetic document images, 903 textlines total.</td>
<td class="num">200</td>
<td class="num">903</td>
<td class="num"><strong>92.80%</strong></td>
<td class="num"><strong>23.24%</strong></td>
<td class="num"><strong>51.63%</strong></td>
<td class="num"><strong>1.33s/page</strong></td>
</tr>
</tbody>
</table>
</div>
<div class="academic-callout callout-info">
<div class="callout-title">🔍 Scientific Analysis: The Empirical Breaking Point</div>
Level 3 and Level 4 results demonstrate that Cham-DBNet handles narrow lines (8–14px) and moderate waviness with near-perfect reliability (99.44%–100% detection rate). The catastrophic drop at Level 5 (61.45%) marks the exact <strong>mathematical breaking point</strong> where interlinear wave amplitude exceeds the interlinear gap (\(\text{Amplitude} > \text{Gap}\)), causing strokes from adjacent lines to intersect on the binarized plane.
</div>
<h3 id="sec6-3">6.3 Model Progression & Empirical Validation: V23 vs V24 vs V25 SOTA</h3>
<p>
Across iterative development cycles, we evaluated the trajectory of the Cham-SVTR architecture across three foundational milestones: <strong>Version 23</strong> (visual ordering legacy), <strong>Version 24</strong> (first logical ordering baseline), and <strong>Version 25</strong> (current production State-Of-The-Art). Quantitative validation on the standardized 10,000 frozen validation split (<code>cham_v25_val_freeze.zip</code>) documents the compounding gains achieved through lexicon expansion, resolution widening, and targeted hard-example synthesis:
</p>
<div class="table-responsive">
<table>
<thead>
<tr>
<th>Metric / Dimension</th>
<th class="num">Version 23 (Visual Legacy)</th>
<th class="num">Version 24 (Logical Baseline)</th>
<th class="num">Version 25 (SOTA Production)</th>
</tr>
</thead>
<tbody>
<tr>
<td><strong>Orthographic Representation</strong></td>
<td class="num">Visual Rendering Order</td>
<td class="num">Logical Brahmic Order</td>
<td class="num"><strong style="color: var(--accent-primary);">Logical Brahmic Order</strong></td>
</tr>
<tr>
<td><strong>Lexicon Dictionary Size</strong></td>
<td class="num">153 tokens</td>
<td class="num">156 tokens</td>
<td class="num"><strong>162 tokens</strong> (+Numerals 1–99, Mark <span class="cham-script">꩞</span>)</td>
</tr>
<tr>
<td><strong>Input Resolution \((C, H, W)\)</strong></td>
<td class="num">\([3, 48, 320]\)</td>
<td class="num">\([3, 48, 320]\)</td>
<td class="num"><strong>\([3, 48, 480]\)</strong> (Long Stanzas)</td>
</tr>
<tr>
<td><strong>Frozen Val Sequence Accuracy</strong></td>
<td class="num">84.12%</td>
<td class="num">88.94%</td>
<td class="num"><strong style="color: var(--color-success, #2e7d32);">90.96%</strong> (+2.02% over V24)</td>
</tr>
<tr>
<td><strong>Normalized Edit Distance (Norm-ED)</strong></td>
<td class="num">96.85%</td>
<td class="num">98.61%</td>
<td class="num"><strong style="color: var(--color-success, #2e7d32);">99.28%</strong></td>
</tr>
<tr>
<td><strong>Character Error Rate (CER)</strong></td>
<td class="num">3.15%</td>
<td class="num">1.39%</td>
<td class="num"><strong style="color: var(--color-success, #2e7d32);">0.72%</strong> (-48.2% vs V24)</td>
</tr>
<tr>
<td><strong>Inference Throughput (Dual T4)</strong></td>
<td class="num">158.2 FPS</td>
<td class="num">154.6 FPS</td>
<td class="num"><strong>143.57 FPS</strong></td>
</tr>
<tr>
<td><strong>Training Regimen</strong></td>
<td class="num">40 epochs (1 stage)</td>
<td class="num">40 epochs (2 stages)</td>
<td class="num"><strong>40 epochs (3 stages, 87,480 steps)</strong></td>
</tr>
<tr>
<td><strong>Model Parameter Footprint</strong></td>
<td class="num">58.82 MB</td>
<td class="num">58.83 MB</td>
<td class="num"><strong>58.84 MB</strong> (<code>best_accuracy.pdparams</code>)</td>
</tr>
</tbody>
</table>
</div>
<h4 style="margin-top: 1.5rem; font-size: 1.05rem; color: var(--text-primary);">Detailed Error Mode Resolution & Paleographic Breakthroughs in V25</h4>
<ol style="margin-bottom: 1.5rem; line-height: 1.7; color: var(--text-secondary);">
<li><strong>Stanza Numerals vs Consonant Disambiguation</strong>:
Prior iterations suffered from CTC blank collapse where Western and Eastern Cham digit ligatures were decoded as visually similar consonants: <span class="cham-script">꩔</span> (digit 4) was misclassified as <span class="cham-script">ꨤ</span> (<em>la</em>), <span class="cham-script">꩕</span> (digit 5) as <span class="cham-script">ꨅ</span> (<em>e</em>) or <span class="cham-script">ꨂ</span> (<em>u</em>), and verse markers like <span class="cham-script">꩑꩞</span> collapsed into <span class="cham-script">ꨩꩌ</span>. Incorporating 18,000 dedicated adversarial samples covering numerals \(1\)–\(99\) with boundary markers (<span class="cham-script">꩑꩞</span> through <span class="cham-script">꩙꩙꩞</span>) achieved \(>98.5\%\) exact match on stanza headers.
</li>
<li><strong>Boundary Punctuation & Double Danda (<span class="cham-script">꩝꩝</span>) Spacing</strong>:
CTC decoders with standard temporal stride tend to merge closely adjacent vertical strokes into a single timestep. V25 synthesized variable Double Danda spacings (\(2\)px to \(8\)px) under varying blur kernels, training the network to distinguish single <span class="cham-script">꩝</span> (U+AA5D) from double <span class="cham-script">꩝꩝</span> (U+AA5D U+AA5D).
</li>
<li><strong>Diacritic Minimal Pair Disambiguation (<span class="cham-script">ꨲ</span> vs <span class="cham-script">ꨶ</span>)</strong>:
Under degraded ink conditions, Vowel Sign UE (<span class="cham-script">ꨲ</span>, U+AA32) and Medial Consonant Sign WA (<span class="cham-script">ꨶ</span>, U+AA36) exhibit extreme visual proximity under base glyphs <span class="cham-script">ꨀ</span>, <span class="cham-script">ꨓ</span>, <span class="cham-script">ꨚ</span>, <span class="cham-script">ꨆ</span>. Through 12,000 targeted adversarial minimal pairs, V25 forces the convolutional feature extractor to attend to subtle sub-stroke curvature differences, dropping the pair substitution rate from \(18.4\%\) in V24 to \(<1.2\%\) in V25.
</li>
<li><strong>Integration with Deterministic Unicode Cluster Normalization</strong>:
Operating strictly in Logical Storage Order, V25 pairs directly with <code>normalize_unicode</code> (via <code>scripts/generate_data.py</code>). Any localized CTC sequence permutations are deterministically regularized to the canonical Brahmic order: \(\text{Base} + \text{Medials} + \text{Pre-Ra} + \text{Pre-Vowels} + \text{Post-Vowels} + \text{Finals}\), eliminating ordering inversion errors while maintaining 100% compliance with modern Unicode display engines.
</li>
</ol>
<h3 id="sec6-4">6.4 Multi-GPU Distributed Hardware Scaling & Cost Efficiency</h3>
<p>
Training deep OCR backbones over 10,000,000 synthetic sample passes requires empirical hardware optimization. We executed controlled distributed benchmarks across four accelerator configurations:
</p>
<div class="table-responsive">
<table>
<thead>
<tr>
<th>Accelerator Environment</th>
<th class="num">Batch Size</th>
<th class="num">Throughput (IPS)</th>
<th class="num">GPU Utilization</th>
<th class="num">VRAM Usage</th>
<th class="num">Full 40 Epochs</th>
<th class="num">Est. Cost ($)</th>
<th class="num">Cost Efficiency (IPS/$)</th>
</tr>
</thead>
<tbody>
<tr>
<td><strong>NVIDIA A100 (80GB SXM4)</strong></td>
<td class="num">192</td>
<td class="num"><strong>255.9</strong></td>
<td class="num">95%</td>
<td class="num">88.5%</td>
<td class="num"><strong>10.85 hrs</strong></td>
<td class="num"><strong>$23.77</strong></td>
<td class="num"><span class="metric-badge metric-good">116.8</span> (Best)</td>
</tr>
<tr>
<td><strong>NVIDIA L40S (48GB Ada)</strong></td>
<td class="num">128</td>
<td class="num"><strong>209.9</strong></td>
<td class="num">94%</td>
<td class="num">52.5%</td>
<td class="num">13.23 hrs</td>
<td class="num">$28.32</td>
<td class="num"><span class="metric-badge metric-good">98.1</span></td>
</tr>
<tr>
<td><strong>Kaggle Dual Tesla T4x2 (DDP)</strong></td>
<td class="num">64 (32x2)</td>
<td class="num"><strong>120.4</strong></td>
<td class="num">92%</td>
<td class="num">83.8%</td>
<td class="num">23.07 hrs</td>
<td class="num"><strong>Free Quota</strong></td>
<td class="num"><span class="metric-badge metric-good">Infinite</span></td>
</tr>
<tr>
<td><strong>NVIDIA L4 (24GB)</strong></td>
<td class="num">96</td>
<td class="num"><strong>74.3</strong></td>
<td class="num">98%</td>
<td class="num">79.2%</td>
<td class="num">37.38 hrs</td>
<td class="num">$29.53</td>
<td class="num">94.0</td>
</tr>
</tbody>
</table>
</div>
</section>
<!-- SECTION 7 -->
<section id="sec7">
<h2>7. Production Serving & Deployment Infrastructure</h2>
<h3 id="sec7-1">7.1 Containerized Microservice Architecture</h3>
<p>
The production Cham OCR Studio is packaged as a lightweight, multi-stage Docker container deployed on <strong>Hugging Face Spaces</strong> (Space ID: <code>phucsd/cham-ocr-studio</code>) and routed through the custom apex domain:
</p>
<div class="math-block">
$$\text{Production URL: } \mathbf{\text{https://ocr.cham.asia}}$$
</div>
<h3 id="sec7-2">7.2 Concurrency Safeguards & NumPy 2.x Forward Compatibility</h3>
<p>
To guarantee deterministic latency under free-tier CPU constraints (2 vCPUs, 16GB RAM):
</p>
<ul>
<li><strong>Single-Thread Concurrency Locks</strong>: OpenMP and MKL thread pools are constrained via <code>OMP_NUM_THREADS=1</code> and <code>CPU_THREADS=1</code>. This eliminates OS context-switching overhead and CPU thrashing during concurrent web requests.</li>
<li><strong>NumPy 2.x Compatibility Layer</strong>: Python 3.12+ environments encounter breaking attribute removals in NumPy 2.x (such as deprecated <code>np.sctypes</code>, <code>np.bool</code>, and <code>np.typeDict</code> required by legacy PaddleOCR and <code>imgaug</code> modules). We injected an automatic runtime patch at module initialization to guarantee zero-crash execution:
<div class="code-box">
<div class="code-header"><span>Python Runtime Monkeypatch</span></div>
<pre><code>import numpy as np
if not hasattr(np, 'sctypes'):
np.sctypes = {'int': [np.int8, np.int16, np.int32, np.int64],
'uint': [np.uint8, np.uint16, np.uint32, np.uint64],
'float': [np.float16, np.float32, np.float64],
'complex': [np.complex64, np.complex128],
'others': [bool, object, bytes, str]}
if not hasattr(np, 'bool'): np.bool = bool
if not hasattr(np, 'int'): np.int = int
if not hasattr(np, 'float'): np.float = float
if not hasattr(np, 'typeDict'): np.typeDict = {}</code></pre>
</div>
</li>
</ul>
</section>
<!-- SECTION 8 -->
<section id="sec8">
<h2>8. Limitations & Open Research Challenges</h2>
<p>
While our pipeline demonstrates substantial gains on synthetic and semi-controlled documents, rigorous academic honesty requires acknowledging the following boundaries:
</p>
<ul>
<li><strong>Synthetic-to-Real Domain Gap</strong>: The neural recognizer is trained predominantly on synthetic textlines rendered with bundled <em>Noto Sans Cham</em> (SIL OFL 1.1). Historical manuscripts feature idiosyncratic scribal hands, variable ink viscosity, and non-standard ligature variations that may exhibit lower recognition confidence in the field.</li>
<li><strong>Absence of Large-Scale Real Manuscript Ground Truth</strong>: Due to the acute scarcity of digitized historical Cham archives with line-level annotations, current benchmarks rely on controlled synthetic stress-tests and curated test slices. Establishing an open, expert-verified historical palm-leaf manuscript benchmark represents an urgent necessity requiring future collaborative scholarship with native Cham elders and linguists.</li>
<li><strong>Severe Biological Degradation & Epigraphy</strong>: Inscription stone rubbings and severely mold-damaged palm leaves (where character ink has substantially eroded) remain beyond the capabilities of pure vision-based CTC sequence models. Integrating masked language models (MLMs) trained on historical Cham corpora is an active subject of future study.</li>
<li><strong>Extreme Interlinear Overlap (Level 5)</strong>: When severe paper wrinkling causes interlinear wave amplitude to exceed line spacing (\(\text{Amplitude} > \text{Gap}\)), 1D projection and horizontal bounding boxes collapse, indicating the need for 2D polygonal baseline tracking models in future iterations.</li>
</ul>
<h3 id="sec8-5">8.5 Post-V25 Future Iterations (Planned V25.1 / V26 Improvements)</h3>
<p>
Following the successful training and deployment of Model Version 25 (90.96% Sequence Accuracy), all subsequent pipeline enhancements are tracked systematically in <a href="https://github.com/phucsd/Cham-OCR/blob/main/FUTURE_WORK.md" target="_blank" rel="noopener noreferrer" style="color: var(--accent-primary); font-weight: 600;">FUTURE_WORK.md</a>:
</p>
<ol style="margin-bottom: 1.5rem; line-height: 1.7; color: var(--text-secondary);">
<li><strong>Dual-Script Font Fallback & Combined Glyph Validation</strong>: Coordinated rendering pipeline for mixed Cham and Vietnamese/Latin text with fontTools cmap validation across both scripts to eliminate tofu artifacts in bilingual annotations.</li>
<li><strong>Dedicated Directional Motion-Blur Augmentation</strong>: Physics-grounded Point Spread Function (PSF) kernels modeling hand tremor motion angles and camera exposure duration.</li>
<li><strong>Manifest-Driven Exact Package Allocation</strong>: Deterministic line count allocation per category instead of probabilistic multinomial sampling.</li>
<li><strong>Deterministic Frozen Validation Generation</strong>: Pre-generated, permanent frozen validation splits for zero-variance benchmark comparisons across training stages.</li>
</ol>
</section>
<!-- SECTION 9 -->
<section id="sec9">
<h2>9. BibTeX Citation & Academic References</h2>
<p>
If you use this work, the synthetic generation methodology, or the OCR Studio in your research, please cite:
</p>
<div class="code-box">
<div class="code-header">
<span>BibTeX Citation</span>
<button class="copy-btn" onclick="navigator.clipboard.writeText(document.getElementById('bibtex-code').innerText)">Copy BibTeX</button>
</div>
<pre id="bibtex-code"><code>@software{cham_ocr_pipeline,
author = {Nguyen, Phuc H.},
title = {Cham-OCR: A Deep-Learning OCR Pipeline for Unicode Eastern Cham, toward Historical Manuscript and Epigraphic Recognition},
year = {2026},
publisher = {GitHub},
url = {https://github.com/phucsd/Cham-OCR},
note = {Production web service: https://ocr.cham.asia}
}</code></pre>
</div>
<h3>Academic References</h3>
<div class="ref-item">
[1] Everson, M. (2006). <em>Proposal for encoding the Cham script in the UCS (ISO/IEC JTC1/SC2/WG2 N3120)</em>. Unicode Consortium. <a href="https://www.unicode.org/L2/L2006/06257-n3120-cham.pdf" target="_blank" rel="noopener noreferrer">https://www.unicode.org/L2/L2006/06257-n3120-cham.pdf</a>
</div>
<div class="ref-item">
[2] Nguyen, M.-T., Schweyer, A.-V., Le, T.-L., Tran, T.-H., & Vu, H. (2019a). <em>Preliminary Results on Ancient Cham Glyph Recognition from Cham Inscription Images</em>. In 2019 International Conference on Multimedia Analysis and Pattern Recognition (MAPR 2019), IEEE, pp. 1–6. DOI: <a href="https://doi.org/10.1109/MAPR.2019.8743540" target="_blank" rel="noopener noreferrer">10.1109/MAPR.2019.8743540</a>
</div>
<div class="ref-item">
[3] Nguyen, M.-T., Schweyer, A.-V., Le, T.-L., Tran, T.-H., & Vu, H. (2019b). <em>Improving Ancient Cham Glyph Recognition Using Data Augmentation and Transfer Learning</em>. In International Conference on Image Analysis and Processing (ICIAP 2019 Workshops: PatReCH), Springer, Cham, LNCS 11808, pp. 115–125. DOI: <a href="https://doi.org/10.1007/978-3-030-30754-7_12" target="_blank" rel="noopener noreferrer">10.1007/978-3-030-30754-7_12</a>
</div>
<div class="ref-item">
[4] Nguyen, T.-N. (2023). <em>Segmentation, Recognition and Indexing of Cham characters in Cham documents</em> (<em>Segmentation, reconnaissance et indexation des caractères Cham dans les documents Cham</em>). Ph.D. Dissertation, Université de La Rochelle, France.
</div>
<div class="ref-item">
[5] Nguyen, T.-N., Burie, J.-C., Le, T.-L., & Schweyer, A.-V. (2023). <em>A Two-Step Sequence Transformer Based Method for Cham to Latin Script Transliteration</em>. In Proceedings of the 7th International Workshop on Historical Document Imaging and Processing (HIP@ICDAR 2023), ACM, pp. 25–30. DOI: <a href="https://doi.org/10.1145/3604951.3605525" target="_blank" rel="noopener noreferrer">10.1145/3604951.3605525</a>
</div>
<div class="ref-item">
[6] Nguyen, T.-N., Burie, J.-C., Le, T.-L., & Schweyer, A.-V. (2022). <em>An effective method for text line segmentation in historical document images</em>. In 26th International Conference on Pattern Recognition (ICPR 2022), IEEE, pp. 2686–2692. DOI: <a href="https://doi.org/10.1109/ICPR56361.2022.9956617" target="_blank" rel="noopener noreferrer">10.1109/ICPR56361.2022.9956617</a>
</div>
<div class="ref-item">
[7] Nguyen, T.-N., Burie, J.-C., Le, T.-L., & Schweyer, A.-V. (2025). <em>Text line segmentation approach combining deep learning model and traditional image processing techniques - application to transliteration of Cham manuscripts</em>. <em>Multimedia Tools and Applications</em>, Springer. DOI: <a href="https://doi.org/10.1007/s11042-025-20602-x" target="_blank" rel="noopener noreferrer">10.1007/s11042-025-20602-x</a>
</div>
<div class="ref-item">
[8] ANR CHAMDOC Project. (2019–2024). <em>Cham Documentation (CHAMDOC)</em>, Project ANR-19-CE27-0018, Agence Nationale de la Recherche, France.
</div>
<div class="ref-item">
[9] Liao, M., Wan, Z., Yao, C., Chen, K., & Bai, X. (2020). <em>Real-time Scene Text Detection with Differentiable Binarization</em>. Proceedings of the AAAI Conference on Artificial Intelligence, 34(07), 11474–11481.
</div>
<div class="ref-item">
[10] Du, Y., Chen, Z., Jia, C., Yin, X., Zheng, T., Li, C., ... & Yu, K. (2023). <em>PP-OCRv4: A Compact, Accurate and Practical Ultra-Lightweight OCR System</em>. arXiv preprint arXiv:2309.09941.
</div>
<div class="ref-item">
[11] Du, Y., Chen, Z., Jia, C., Yin, X., Zheng, T., Li, C., ... & Yu, K. (2022). <em>SVTR: Scene Text Recognition with a Single Visual Model</em>. arXiv preprint arXiv:2205.00159.
</div>
<div class="ref-item">
[12] Graves, A., Fernández, S., Gomez, F., & Schmidhuber, J. (2006). <em>Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks</em>. Proceedings of the 23rd International Conference on Machine Learning (ICML '06), 369–376.
</div>
</section>
<!-- Footer -->
<footer class="paper-footer">
<div>© 2026 Phuc H. Nguyen. Released under the MIT License. | Contact: <a href="mailto:phucsd@gmail.com" style="color: var(--accent-primary); font-weight: 600;">phucsd@gmail.com</a></div>
<div>Official Research Repository: <a href="https://github.com/phucsd/Cham-OCR" target="_blank" rel="noopener noreferrer" style="color: var(--accent-primary); font-weight: 600;">github.com/phucsd/Cham-OCR</a></div>
</footer>
</article>
</div>
<!-- Active Table of Contents ScrollSpy & Anchor Offset Script -->
<script>
document.addEventListener('DOMContentLoaded', () => {
const links = document.querySelectorAll('.toc-link, .toc-sublink');
const sections = [];
const NAV_OFFSET = 85; // 58px navbar + 27px breathing margin
links.forEach(link => {
const href = link.getAttribute('href');
if (href && href.startsWith('#')) {
const target = document.querySelector(href);
if (target) {
sections.push({ link, target });
link.addEventListener('click', (e) => {
e.preventDefault();
const targetTop = target.getBoundingClientRect().top + window.pageYOffset - NAV_OFFSET;
window.scrollTo({
top: targetTop,
behavior: 'smooth'
});
history.pushState(null, null, href);
});
}
}
});
// Handle direct page load with hash in URL
if (window.location.hash) {
setTimeout(() => {
const target = document.querySelector(window.location.hash);
if (target) {
const targetTop = target.getBoundingClientRect().top + window.pageYOffset - NAV_OFFSET;
window.scrollTo({
top: targetTop,
behavior: 'smooth'
});
}
}, 150);
}
// ScrollSpy
function onScroll() {
const scrollPos = window.scrollY + 120;
for (let i = sections.length - 1; i >= 0; i--) {
const sec = sections[i];
const elementTop = sec.target.getBoundingClientRect().top + window.pageYOffset;
if (elementTop <= scrollPos) {
links.forEach(l => l.classList.remove('active'));
sec.link.classList.add('active');
break;
}
}
}
window.addEventListener('scroll', onScroll, { passive: true });
onScroll();
});
</script>
</body>
</html>
|