RavindranadhM commited on
Commit
1925f28
·
verified ·
1 Parent(s): a3abb2d

Add trained steel ROI gate and hybrid validation

Browse files
inspection/surface_gate.py CHANGED
@@ -188,18 +188,43 @@ class SteelSurfaceGate:
188
  roi_area_ratio=roi_area_ratio,
189
  )
190
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
191
  return SurfaceGateResult(
192
  is_steel=True,
193
  confidence=float(best_candidate["confidence"]),
194
- mode="detector",
195
  label=str(best_candidate["label"]),
196
  reason=(
197
- f"Steel-surface detector localized an inspection ROI with {float(best_candidate['confidence']):.0%} confidence."
 
198
  ),
199
  metrics={
200
  "boxes_detected": float(len(candidates)),
201
  "roi_area_ratio": round(roi_area_ratio, 4),
202
  "selector_score": round(float(best_candidate["selector_score"]), 4),
 
203
  },
204
  roi_bbox=best_candidate["bbox"],
205
  roi_area_ratio=roi_area_ratio,
 
188
  roi_area_ratio=roi_area_ratio,
189
  )
190
 
191
+ roi_x, roi_y, roi_width, roi_height = best_candidate["bbox"]
192
+ roi_image = image[roi_y:roi_y + roi_height, roi_x:roi_x + roi_width]
193
+ heuristic_result = self._evaluate_with_heuristic(roi_image)
194
+ if not heuristic_result.is_steel:
195
+ return SurfaceGateResult(
196
+ is_steel=False,
197
+ confidence=float(best_candidate["confidence"]),
198
+ mode="detector+heuristic",
199
+ label=str(best_candidate["label"]),
200
+ reason=(
201
+ "Frame skipped because the localized ROI did not pass the steel-surface texture validation step. "
202
+ "Reduce background content and center the actual material."
203
+ ),
204
+ metrics={
205
+ "boxes_detected": float(len(candidates)),
206
+ "roi_area_ratio": round(roi_area_ratio, 4),
207
+ "selector_score": round(float(best_candidate["selector_score"]), 4),
208
+ **heuristic_result.metrics,
209
+ },
210
+ roi_bbox=best_candidate["bbox"],
211
+ roi_area_ratio=roi_area_ratio,
212
+ )
213
+
214
  return SurfaceGateResult(
215
  is_steel=True,
216
  confidence=float(best_candidate["confidence"]),
217
+ mode="detector+heuristic",
218
  label=str(best_candidate["label"]),
219
  reason=(
220
+ f"Steel-surface detector localized an inspection ROI with {float(best_candidate['confidence']):.0%} confidence, "
221
+ "and the ROI passed surface-texture validation."
222
  ),
223
  metrics={
224
  "boxes_detected": float(len(candidates)),
225
  "roi_area_ratio": round(roi_area_ratio, 4),
226
  "selector_score": round(float(best_candidate["selector_score"]), 4),
227
+ **heuristic_result.metrics,
228
  },
229
  roi_bbox=best_candidate["bbox"],
230
  roi_area_ratio=roi_area_ratio,
models/steel_surface_detector.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fcf2d6d9662e8ee5765db2d27f313b4155e3688ea946f335af9dd01e4fff91b4
3
+ size 5436442
training/README.md CHANGED
@@ -34,7 +34,9 @@ Positive steel sources:
34
 
35
  Automatic non-steel/background source:
36
 
37
- - Imagenette via Hugging Face Datasets
 
 
38
 
39
  This avoids making you manually collect a non-steel dataset.
40
 
@@ -54,7 +56,13 @@ cd "/Users/ravindranadhm/Documents/Projects/steel-surface-inspection copy"
54
  venv/bin/python training/prepare_surface_gate_dataset.py
55
  ```
56
 
57
- This default mode is the fastest path because it avoids downloading the larger Severstal archive.
 
 
 
 
 
 
58
 
59
  If you want a stronger detector after the first run, rebuild with Severstal included:
60
 
 
34
 
35
  Automatic non-steel/background source:
36
 
37
+ - `torchvision` datasets downloaded automatically
38
+ - default quick path: `CIFAR10`
39
+ - optional stronger but heavier path: `STL10`
40
 
41
  This avoids making you manually collect a non-steel dataset.
42
 
 
56
  venv/bin/python training/prepare_surface_gate_dataset.py
57
  ```
58
 
59
+ This default mode is the fastest path because it avoids the larger Severstal archive and uses `CIFAR10` as an automatic non-steel source.
60
+
61
+ If you want stronger natural-scene negatives later:
62
+
63
+ ```bash
64
+ venv/bin/python training/prepare_surface_gate_dataset.py --negative-source stl10
65
+ ```
66
 
67
  If you want a stronger detector after the first run, rebuild with Severstal included:
68
 
training/prepare_surface_gate_dataset.py CHANGED
@@ -3,7 +3,6 @@ from __future__ import annotations
3
  import argparse
4
  import json
5
  import random
6
- import shutil
7
  import zipfile
8
  from pathlib import Path
9
 
@@ -12,18 +11,11 @@ import numpy as np
12
  import yaml
13
  from huggingface_hub import snapshot_download
14
  from tqdm import tqdm
15
-
16
- try:
17
- from datasets import load_dataset
18
- except Exception as exc: # pragma: no cover - import guard for guidance
19
- raise SystemExit(
20
- "The `datasets` package is required for this script. Install training/requirements.txt first."
21
- ) from exc
22
 
23
 
24
  SEVERSTAL_REPO = "rohanath/severstal-steel-detection"
25
  NEU_REPO = "LiuErXiao/NEU_valid"
26
- IMAGENETTE_REPO = "frgfm/imagenette"
27
 
28
 
29
  def parse_args() -> argparse.Namespace:
@@ -76,6 +68,12 @@ def parse_args() -> argparse.Namespace:
76
  default=1.0,
77
  help="How many pure background negatives to generate per steel image.",
78
  )
 
 
 
 
 
 
79
  parser.add_argument("--seed", type=int, default=42)
80
  return parser.parse_args()
81
 
@@ -187,19 +185,44 @@ def save_image(path: Path, image: np.ndarray) -> None:
187
  cv2.imwrite(str(path), image)
188
 
189
 
190
- def export_negative_pool(target_dir: Path, limit: int, rng: random.Random) -> list[Path]:
191
- dataset = load_dataset(IMAGENETTE_REPO, "160px")
192
- samples = list(dataset["train"]) + list(dataset["validation"])
193
- rng.shuffle(samples)
194
-
195
- output_paths: list[Path] = []
196
- for index, sample in enumerate(samples[:limit]):
197
- image = cv2.cvtColor(np.array(sample["image"]), cv2.COLOR_RGB2BGR)
198
- destination = target_dir / f"imagenette_{index:05d}.jpg"
199
- save_image(destination, center_crop_and_resize(image, 640))
200
- output_paths.append(destination)
201
-
202
- return output_paths
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
203
 
204
 
205
  def build_split(
@@ -293,7 +316,7 @@ def main() -> None:
293
  ensure_clean_dir(labels_dir)
294
 
295
  neu_raw = extract_first_zip(NEU_REPO, raw_root / "neu")
296
- negative_dir = raw_root / "imagenette"
297
  ensure_clean_dir(negative_dir)
298
 
299
  steel_paths: list[Path] = []
@@ -309,7 +332,12 @@ def main() -> None:
309
  if not steel_paths:
310
  raise RuntimeError("No steel images were collected for the detector dataset.")
311
 
312
- negative_paths = export_negative_pool(negative_dir, args.max_negatives, rng)
 
 
 
 
 
313
  if not negative_paths:
314
  raise RuntimeError("No negative background images were collected.")
315
 
@@ -343,6 +371,7 @@ def main() -> None:
343
  {
344
  "steel_sources": len(steel_paths),
345
  "negative_pool": len(negative_paths),
 
346
  "splits": summary,
347
  "dataset_yaml": str(yaml_path),
348
  },
 
3
  import argparse
4
  import json
5
  import random
 
6
  import zipfile
7
  from pathlib import Path
8
 
 
11
  import yaml
12
  from huggingface_hub import snapshot_download
13
  from tqdm import tqdm
14
+ from torchvision.datasets import CIFAR10, STL10
 
 
 
 
 
 
15
 
16
 
17
  SEVERSTAL_REPO = "rohanath/severstal-steel-detection"
18
  NEU_REPO = "LiuErXiao/NEU_valid"
 
19
 
20
 
21
  def parse_args() -> argparse.Namespace:
 
68
  default=1.0,
69
  help="How many pure background negatives to generate per steel image.",
70
  )
71
+ parser.add_argument(
72
+ "--negative-source",
73
+ default="cifar10",
74
+ choices=["cifar10", "stl10", "auto"],
75
+ help="Automatic non-steel image source. `cifar10` is the fastest default.",
76
+ )
77
  parser.add_argument("--seed", type=int, default=42)
78
  return parser.parse_args()
79
 
 
185
  cv2.imwrite(str(path), image)
186
 
187
 
188
+ def export_negative_pool(
189
+ target_dir: Path,
190
+ limit: int,
191
+ rng: random.Random,
192
+ negative_source: str,
193
+ ) -> tuple[list[Path], str]:
194
+ source_loaders = {
195
+ "cifar10": lambda root: CIFAR10(root=str(root), train=True, download=True),
196
+ "stl10": lambda root: STL10(root=str(root), split="train", download=True),
197
+ }
198
+ if negative_source == "auto":
199
+ datasets_to_try = [("cifar10", source_loaders["cifar10"]), ("stl10", source_loaders["stl10"])]
200
+ else:
201
+ datasets_to_try = [(negative_source, source_loaders[negative_source])]
202
+
203
+ last_error: Exception | None = None
204
+ for dataset_name, loader in datasets_to_try:
205
+ try:
206
+ dataset = loader(target_dir / "_torchvision_cache")
207
+ samples = list(range(len(dataset)))
208
+ rng.shuffle(samples)
209
+
210
+ output_paths: list[Path] = []
211
+ for index, sample_index in enumerate(samples[:limit]):
212
+ sample_image, _ = dataset[sample_index]
213
+ image = cv2.cvtColor(np.array(sample_image), cv2.COLOR_RGB2BGR)
214
+ destination = target_dir / f"{dataset_name}_{index:05d}.jpg"
215
+ save_image(destination, center_crop_and_resize(image, 640))
216
+ output_paths.append(destination)
217
+
218
+ if output_paths:
219
+ return output_paths, dataset_name
220
+ except Exception as exc: # pragma: no cover - download/runtime fallback
221
+ last_error = exc
222
+
223
+ raise RuntimeError(
224
+ "Unable to download an automatic negative-image source via torchvision."
225
+ ) from last_error
226
 
227
 
228
  def build_split(
 
316
  ensure_clean_dir(labels_dir)
317
 
318
  neu_raw = extract_first_zip(NEU_REPO, raw_root / "neu")
319
+ negative_dir = raw_root / "negatives"
320
  ensure_clean_dir(negative_dir)
321
 
322
  steel_paths: list[Path] = []
 
332
  if not steel_paths:
333
  raise RuntimeError("No steel images were collected for the detector dataset.")
334
 
335
+ negative_paths, negative_source = export_negative_pool(
336
+ negative_dir,
337
+ args.max_negatives,
338
+ rng,
339
+ args.negative_source,
340
+ )
341
  if not negative_paths:
342
  raise RuntimeError("No negative background images were collected.")
343
 
 
371
  {
372
  "steel_sources": len(steel_paths),
373
  "negative_pool": len(negative_paths),
374
+ "negative_source": negative_source,
375
  "splits": summary,
376
  "dataset_yaml": str(yaml_path),
377
  },
training/requirements.txt CHANGED
@@ -1,4 +1,3 @@
1
- datasets
2
  huggingface_hub
3
  Pillow
4
  PyYAML
 
 
1
  huggingface_hub
2
  Pillow
3
  PyYAML
training/train_surface_gate_detector.py CHANGED
@@ -23,7 +23,7 @@ def parse_args() -> argparse.Namespace:
23
  )
24
  parser.add_argument("--epochs", type=int, default=18)
25
  parser.add_argument("--imgsz", type=int, default=512)
26
- parser.add_argument("--batch", default=16)
27
  parser.add_argument("--device", default="")
28
  parser.add_argument("--workers", type=int, default=4)
29
  parser.add_argument("--project", default="training/runs")
@@ -42,12 +42,14 @@ def main() -> None:
42
  if not data_path.exists():
43
  raise FileNotFoundError(f"Dataset YAML not found: {data_path}")
44
 
 
 
45
  model = YOLO(args.model)
46
  results = model.train(
47
  data=str(data_path),
48
  epochs=args.epochs,
49
  imgsz=args.imgsz,
50
- batch=args.batch,
51
  device=args.device or None,
52
  workers=args.workers,
53
  project=args.project,
 
23
  )
24
  parser.add_argument("--epochs", type=int, default=18)
25
  parser.add_argument("--imgsz", type=int, default=512)
26
+ parser.add_argument("--batch", type=float, default=16)
27
  parser.add_argument("--device", default="")
28
  parser.add_argument("--workers", type=int, default=4)
29
  parser.add_argument("--project", default="training/runs")
 
42
  if not data_path.exists():
43
  raise FileNotFoundError(f"Dataset YAML not found: {data_path}")
44
 
45
+ batch = int(args.batch) if float(args.batch).is_integer() else float(args.batch)
46
+
47
  model = YOLO(args.model)
48
  results = model.train(
49
  data=str(data_path),
50
  epochs=args.epochs,
51
  imgsz=args.imgsz,
52
+ batch=batch,
53
  device=args.device or None,
54
  workers=args.workers,
55
  project=args.project,
yolo11n.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0ebbc80d4a7680d14987a577cd21342b65ecfd94632bd9a8da63ae6417644ee1
3
+ size 5613764