File size: 8,062 Bytes
d667566 | 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 | """Synthetic pair generator.
Reproduces the dataset's own construction: take a clean template, paste
differences into a copy of it, then degrade that copy into a "photo" (blur,
noise, JPEG, tone curve, shadow field, sub-pixel warp). The generated pairs are
written through the same preprocessing path as the real ones, so training code
cannot tell them apart.
Edit types and sizes follow the measured training statistics: median 22x22 px,
30% of boxes under 16 px, additions and modifications only, never deletions.
"""
from __future__ import annotations
import json
from pathlib import Path
import cv2
import numpy as np
from .config import DATA, SYNTH
from .preprocess import local_norm, match_blur
EDIT_WEIGHTS = {"swap": 0.40, "insert": 0.30, "mark": 0.30}
def ink_components(img: np.ndarray, min_area: int = 12, max_area: int = 8000):
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
binary = cv2.adaptiveThreshold(gray, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C,
cv2.THRESH_BINARY_INV, 31, 12)
n, _, stats, _ = cv2.connectedComponentsWithStats(binary, 8)
out = []
for i in range(1, n):
x, y, w, h, area = stats[i]
if min_area <= area <= max_area and 3 <= w <= 160 and 3 <= h <= 160:
out.append((int(x), int(y), int(w), int(h)))
return out
def blank_spot(ink: np.ndarray, size: int, rng: np.random.RandomState, tries: int = 40):
"""Find a location whose neighbourhood is empty but which sits near content."""
h, w = ink.shape
for _ in range(tries):
x = rng.randint(4, max(5, w - size - 4))
y = rng.randint(4, max(5, h - size - 4))
patch = ink[y - 2:y + size + 2, x - 2:x + size + 2]
if patch.size and patch.max() < 8:
near = ink[max(0, y - 60):y + size + 60, max(0, x - 60):x + size + 60]
if near.size and near.max() > 30:
return x, y
return None
def apply_edits(template: np.ndarray, rng: np.random.RandomState,
n_edits: tuple[int, int] = (3, 11)):
"""Return (edited image, boxes). Edits only add or change ink, never remove it."""
edited = template.copy()
ink = local_norm(template)
comps = ink_components(template)
boxes: list[list[int]] = []
n = rng.randint(*n_edits)
kinds = list(EDIT_WEIGHTS)
probs = np.array([EDIT_WEIGHTS[k] for k in kinds], np.float32)
probs /= probs.sum()
for _ in range(n):
kind = kinds[int(rng.choice(len(kinds), p=probs))]
if kind == "swap" and len(comps) > 2:
x, y, w, h = comps[rng.randint(len(comps))]
for _ in range(10):
sx, sy, sw, sh = comps[rng.randint(len(comps))]
if abs(sw - w) <= max(4, w // 2) and abs(sh - h) <= max(4, h // 2):
break
else:
continue
src = template[sy:sy + sh, sx:sx + sw]
src = cv2.resize(src, (w, h), interpolation=cv2.INTER_AREA)
edited[y:y + h, x:x + w] = src
boxes.append([x, y, x + w, y + h])
elif kind == "insert" and comps:
sx, sy, sw, sh = comps[rng.randint(len(comps))]
spot = blank_spot(ink, max(sw, sh), rng)
if spot is None:
continue
x, y = spot
patch = template[sy:sy + sh, sx:sx + sw]
region = edited[y:y + sh, x:x + sw]
if region.shape != patch.shape:
continue
edited[y:y + sh, x:x + sw] = np.minimum(region, patch)
boxes.append([x, y, x + sw, y + sh])
else: # mark
size = int(rng.choice([6, 8, 8, 10, 12, 16, 24]))
spot = blank_spot(ink, size, rng)
if spot is None:
continue
x, y = spot
colour = int(rng.randint(0, 90))
if rng.rand() < 0.7:
cv2.rectangle(edited, (x, y), (x + size - 1, y + size - 1),
(colour, colour, colour), -1)
else:
cv2.circle(edited, (x + size // 2, y + size // 2), size // 2,
(colour, colour, colour), -1)
boxes.append([x, y, x + size, y + size])
return edited, np.asarray(boxes, np.float32) if boxes else np.zeros((0, 4), np.float32)
def shadow_field(shape: tuple[int, int], rng: np.random.RandomState) -> np.ndarray:
"""Smooth multiplicative illumination field with occasional dark patches."""
h, w = shape
small = rng.uniform(0.75, 1.05, (rng.randint(3, 7), rng.randint(3, 7))).astype(np.float32)
field = cv2.resize(small, (w, h), interpolation=cv2.INTER_CUBIC)
if rng.rand() < 0.35:
mask = np.zeros((h, w), np.float32)
cx, cy = rng.randint(0, w), rng.randint(0, h)
axes = (rng.randint(w // 6, max(w // 6 + 1, w // 2)),
rng.randint(h // 6, max(h // 6 + 1, h // 2)))
cv2.ellipse(mask, (cx, cy), axes, rng.randint(0, 180), 0, 360, 1.0, -1)
mask = cv2.GaussianBlur(mask, (0, 0), max(w, h) / 30.0)
field = field * (1.0 - mask * rng.uniform(0.15, 0.55))
return np.clip(field, 0.2, 1.2)
def degrade(img: np.ndarray, rng: np.random.RandomState) -> np.ndarray:
out = img.astype(np.float32)
if rng.rand() < 0.5: # sub-pixel warp
dx, dy = rng.uniform(-0.3, 0.3), rng.uniform(-0.3, 0.3)
m = np.float32([[1, 0, dx], [0, 1, dy]])
out = cv2.warpAffine(out, m, (out.shape[1], out.shape[0]),
flags=cv2.INTER_LINEAR, borderMode=cv2.BORDER_REPLICATE)
out = cv2.GaussianBlur(out, (0, 0), rng.uniform(0.6, 2.2))
out = out * shadow_field(out.shape[:2], rng)[..., None]
gamma = rng.uniform(0.8, 1.25)
out = 255.0 * np.power(np.clip(out, 0, 255) / 255.0, gamma)
out = out + rng.normal(0, rng.uniform(1.5, 7.0), out.shape)
out = np.clip(out, 0, 255).astype(np.uint8)
quality = int(rng.randint(55, 96))
ok, buf = cv2.imencode(".jpg", out, [int(cv2.IMWRITE_JPEG_QUALITY), quality])
return cv2.imdecode(buf, cv2.IMREAD_COLOR) if ok else out
def generate(n: int, seed: int = 0, out_root: Path | None = None,
source_split: str = "train", max_side: int = 2400) -> dict:
"""Write n synthetic pairs in prepared form. Returns {index: boxes}."""
out_root = Path(out_root or SYNTH) / "synth"
out_root.mkdir(parents=True, exist_ok=True)
templates = sorted((DATA / source_split / "template").glob("*.png"))
if not templates:
raise FileNotFoundError(f"no templates under {DATA / source_split / 'template'}")
boxes_by_index: dict[int, list] = {}
for i in range(n):
rng = np.random.RandomState(seed * 100003 + i)
src = cv2.imread(str(templates[rng.randint(len(templates))]), cv2.IMREAD_COLOR)
if src is None:
continue
if max(src.shape[:2]) > max_side:
scale = max_side / max(src.shape[:2])
src = cv2.resize(src, None, fx=scale, fy=scale, interpolation=cv2.INTER_AREA)
edited, boxes = apply_edits(src, rng)
if len(boxes) == 0:
continue
photo = degrade(edited, rng)
template_matched, _ = match_blur(src, photo)
index = seed * 1000000 + i
d = out_root / f"{index:03d}"
d.mkdir(parents=True, exist_ok=True)
cv2.imwrite(str(d / "t.png"), template_matched)
cv2.imwrite(str(d / "p.png"), photo)
cv2.imwrite(str(d / "tn.png"), local_norm(template_matched))
cv2.imwrite(str(d / "pn.png"), local_norm(photo))
boxes_by_index[index] = boxes.tolist()
return boxes_by_index
def merge_box_index(out_root: Path | None = None) -> int:
"""Merge per-shard box files into the single boxes.json the loader reads."""
root = Path(out_root or SYNTH) / "synth"
merged: dict[str, list] = {}
for shard in sorted(root.glob("boxes_shard_*.json")):
merged.update(json.loads(shard.read_text()))
(root / "boxes.json").write_text(json.dumps(merged))
return len(merged)
|