File size: 17,469 Bytes
7111e1a | 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 | """
dataset.py
==========
PyTorch Dataset and DataLoader utilities for the Civil Registry OCR system.
"""
import os
import json
import random
from pathlib import Path
from typing import List, Tuple, Dict, Optional
import cv2
import numpy as np
import torch
from torch.utils.data import Dataset
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# CHARACTER SET
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
PRINTABLE_CHARS = [chr(i) for i in range(32, 127)] # space (32) to ~ (126)
def build_char_maps(extra_chars: Optional[List[str]] = None):
chars = PRINTABLE_CHARS.copy()
if extra_chars:
for c in extra_chars:
if c not in chars:
chars.append(c)
char_to_idx = {c: i + 1 for i, c in enumerate(chars)}
idx_to_char = {i + 1: c for i, c in enumerate(chars)}
num_chars = len(chars) + 1 # +1 for blank=0
return char_to_idx, idx_to_char, num_chars
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# IMAGE NORMALIZER
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class ImageNormalizer:
def __init__(self, target_height: int = 64, target_width: int = 512):
self.H = target_height
self.W = target_width
def _to_gray(self, img):
if len(img.shape) == 3:
return cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
return img.copy()
def _crop_to_text(self, gray):
inv = cv2.bitwise_not(gray)
_, thresh = cv2.threshold(inv, 20, 255, cv2.THRESH_BINARY)
coords = np.column_stack(np.where(thresh > 0))
if len(coords) == 0:
return gray
y_min, x_min = coords.min(axis=0)
y_max, x_max = coords.max(axis=0)
pad = max(4, int((y_max - y_min) * 0.15))
y_min = max(0, y_min - pad)
x_min = max(0, x_min - pad)
y_max = min(gray.shape[0] - 1, y_max + pad)
x_max = min(gray.shape[1] - 1, x_max + pad)
return gray[y_min:y_max + 1, x_min:x_max + 1]
def _aspect_resize(self, gray):
h, w = gray.shape
if h == 0 or w == 0:
return np.ones((self.H, self.W), dtype=np.uint8) * 255
scale = self.H / h
new_w = int(w * scale)
new_h = self.H
if new_w > self.W:
scale = self.W / w
new_h = int(h * scale)
new_w = self.W
resized = cv2.resize(gray, (new_w, new_h), interpolation=cv2.INTER_LANCZOS4)
canvas = np.ones((self.H, self.W), dtype=np.uint8) * 255
y_off = (self.H - new_h) // 2
x_off = (self.W - new_w) // 2
canvas[y_off:y_off + new_h, x_off:x_off + new_w] = resized
return canvas
def _binarize(self, img):
_, otsu = cv2.threshold(img, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
white_ratio = np.mean(otsu == 255)
if white_ratio < 0.30 or white_ratio > 0.97:
return cv2.adaptiveThreshold(
img, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C,
cv2.THRESH_BINARY, 11, 2)
return otsu
def normalize(self, img: np.ndarray, augmenter=None) -> np.ndarray:
gray = self._to_gray(img)
# NOTE: fastNlMeansDenoising intentionally removed from training pipeline.
# It is slow (~200ms/image) and pointless on clean synthetic images.
# Denoising is only applied in check_cer.py / inference.py (AdaptiveNormalizer)
# which runs on real scanned documents where denoising actually helps.
gray = self._crop_to_text(gray)
gray = self._aspect_resize(gray)
# FIXED Bug 3: augment on grayscale BEFORE binarize.
# Brightness/contrast augmentation has zero effect on binary (0/255) pixels.
if augmenter is not None:
gray = augmenter(gray)
return self._binarize(gray)
def to_tensor(self, img: np.ndarray) -> torch.Tensor:
return torch.FloatTensor(
img.astype(np.float32) / 255.0
).unsqueeze(0) # [1, H, W]
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# AUGMENTATION
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class Augmenter:
def __call__(self, img: np.ndarray) -> np.ndarray:
img = img.copy()
# Random slight rotation (Β±3Β°)
if random.random() < 0.3:
angle = random.uniform(-3, 3)
h, w = img.shape
M = cv2.getRotationMatrix2D((w // 2, h // 2), angle, 1.0)
img = cv2.warpAffine(img, M, (w, h),
borderMode=cv2.BORDER_CONSTANT,
borderValue=255)
# Random brightness/contrast
if random.random() < 0.4:
alpha = random.uniform(0.8, 1.2)
beta = random.randint(-20, 20)
img = np.clip(alpha * img.astype(np.float32) + beta,
0, 255).astype(np.uint8)
# Gaussian blur
if random.random() < 0.3:
ksize = random.choice([3, 5])
img = cv2.GaussianBlur(img, (ksize, ksize), 0)
# Salt-and-pepper noise
if random.random() < 0.2:
noise = np.random.randint(0, 100, img.shape)
img[noise < 2] = 0
img[noise > 97] = 255
# Random small horizontal shift
if random.random() < 0.2:
h, w = img.shape
shift = random.randint(-int(w * 0.05), int(w * 0.05))
M = np.float32([[1, 0, shift], [0, 1, 0]])
img = cv2.warpAffine(img, M, (w, h),
borderMode=cv2.BORDER_CONSTANT,
borderValue=255)
# ββ NEW: Horizontal line noise ββββββββββββββββββββββββββββββββββββββββ
# Simulates ruled form lines bleeding through behind the text.
# Civil registry forms have printed horizontal grid lines β scanners
# often pick these up as faint grey stripes across text fields.
if random.random() < 0.3:
h, w = img.shape
n_lines = random.randint(1, 3)
for _ in range(n_lines):
y = random.randint(0, h - 1)
thickness = random.choice([1, 1, 1, 2]) # mostly 1px
intensity = random.randint(160, 220) # light grey, not black
cv2.line(img, (0, y), (w, y),
color=intensity, thickness=thickness)
# ββ NEW: Perspective warp βββββββββββββββββββββββββββββββββββββββββββββ
# Simulates documents scanned or photographed at a slight angle.
# Keystone distortion is common when forms are placed unevenly on
# a flatbed scanner or photographed with a phone camera.
if random.random() < 0.25:
h, w = img.shape
d = 0.03
dx = int(w * d)
dy = int(h * d)
src = np.float32([[0, 0], [w, 0], [w, h], [0, h]])
dst = np.float32([
[random.randint(0, dx), random.randint(0, dy)],
[w - random.randint(0, dx), random.randint(0, dy)],
[w - random.randint(0, dx), h - random.randint(0, dy)],
[random.randint(0, dx), h - random.randint(0, dy)],
])
M = cv2.getPerspectiveTransform(src, dst)
img = cv2.warpPerspective(img, M, (w, h),
borderMode=cv2.BORDER_CONSTANT,
borderValue=255)
return img
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# DATASET
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
class CivilRegistryDataset(Dataset):
"""
Args:
data_dir : root folder containing image subfolders (e.g. 'data/train')
annotations_file : path to JSON file with image_path + text pairs
img_height : target image height (default 64)
img_width : target image width (default 512)
augment : True = apply augmentation (training only)
form_type : 'all' or filter by form e.g. 'form1a'
Properties used by train.py:
.num_chars β passed to CRNN model
.char_to_idx β saved in checkpoint
.idx_to_char β used for decoding predictions
__getitem__ returns:
image_tensor FloatTensor [1, H, W]
target LongTensor [label_length]
target_length int
text str (original ground truth)
"""
def __init__(
self,
data_dir: str,
annotations_file: str,
img_height: int = 64,
img_width: int = 512,
augment: bool = False,
form_type: str = 'all',
seed: Optional[int] = None, # Rec 2: reproducible augmentation
):
self.data_dir = Path(data_dir)
self.augment = augment
self.normalizer = ImageNormalizer(img_height, img_width)
self.augmenter = Augmenter()
if seed is not None: # Rec 2: seed random for reproducibility
random.seed(seed)
np.random.seed(seed)
self.char_to_idx, self.idx_to_char, self.num_chars = build_char_maps()
with open(annotations_file, 'r', encoding='utf-8') as f:
all_annotations = json.load(f)
if form_type != 'all':
all_annotations = [
a for a in all_annotations
if form_type in a.get('image_path', '')
]
self.samples: List[Dict] = []
missing = 0
for ann in all_annotations:
img_path = self.data_dir / ann['image_path']
if img_path.exists():
text = ann['text'].strip()
if text:
self.samples.append({
'image_path': str(img_path),
'text': text,
})
else:
missing += 1
if missing > 0:
print(f" [Dataset] WARNING: {missing} image(s) not found and skipped.")
print(f" [Dataset] Loaded {len(self.samples)} samples "
f"from {annotations_file} (augment={augment})")
def __len__(self) -> int:
return len(self.samples)
def __getitem__(self, idx: int):
sample = self.samples[idx]
text = sample['text']
img = cv2.imread(sample['image_path'])
if img is None:
img = np.ones((64, 512, 3), dtype=np.uint8) * 255
# FIXED Bug 3: pass augmenter into normalize() so it runs on grayscale
# (before binarization), not on the binary output where it has no effect.
aug = self.augmenter if self.augment else None
normalized = self.normalizer.normalize(img, augmenter=aug)
image_tensor = self.normalizer.to_tensor(normalized) # [1, H, W]
encoded = [
self.char_to_idx[c]
for c in text
if c in self.char_to_idx
]
if len(encoded) == 0:
encoded = [self.char_to_idx.get(' ', 1)]
target = torch.LongTensor(encoded)
target_length = len(encoded)
return image_tensor, target, target_length, text
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# COLLATE FUNCTION
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def collate_fn(batch):
"""
CTC loss needs all labels packed into one flat 1D tensor.
PyTorch's default collator can't handle variable-length labels,
so this custom function packs them correctly.
Returns:
images FloatTensor [B, 1, H, W]
targets LongTensor [sum of all label lengths]
target_lengths LongTensor [B]
texts List[str]
"""
images, targets, target_lengths, texts = zip(*batch)
images = torch.stack(images, dim=0)
targets = torch.cat([t for t in targets])
target_lengths = torch.LongTensor(target_lengths)
return images, targets, target_lengths, list(texts)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# HELPER: CREATE ANNOTATION FILE (run once to build your JSON)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def create_annotation_file(data_dir: str, output_file: str,
extensions=('.jpg', '.jpeg', '.png')):
"""
Auto-generate annotations JSON by scanning data_dir.
For each image, looks for a sidecar .txt file with the same name.
If not found, uses the filename stem (underscores β spaces) as label.
Usage:
from dataset import create_annotation_file
create_annotation_file('data/train', 'data/train_annotations.json')
create_annotation_file('data/val', 'data/val_annotations.json')
"""
data_path = Path(data_dir)
annotations = []
for img_path in sorted(data_path.rglob('*')):
if img_path.suffix.lower() not in extensions:
continue
txt_path = img_path.with_suffix('.txt')
if txt_path.exists():
label = txt_path.read_text(encoding='utf-8').strip()
else:
label = img_path.stem.replace('_', ' ')
if not label:
continue
rel_path = img_path.relative_to(data_path)
annotations.append({
'image_path': str(rel_path).replace('\\', '/'),
'text': label,
})
os.makedirs(Path(output_file).parent, exist_ok=True)
with open(output_file, 'w', encoding='utf-8') as f:
json.dump(annotations, f, indent=2, ensure_ascii=False)
print(f"β Saved {len(annotations)} entries β {output_file}")
return annotations
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
# SELF-TEST (python dataset.py)
# βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
if __name__ == '__main__':
print("=" * 55)
print(" dataset.py self-test")
print("=" * 55)
c2i, i2c, n = build_char_maps()
print(f"\n Vocab size : {n} (including blank=0)")
print(f" 'A'={c2i['A']} '0'={c2i['0']} ' '={c2i[' ']} '.'={c2i['.']}")
dummy = np.ones((80, 300, 3), dtype=np.uint8) * 200
norm = ImageNormalizer(64, 512)
out = norm.normalize(dummy)
t = norm.to_tensor(out)
print(f"\n Normalizer : {dummy.shape} β {out.shape} β tensor {t.shape}")
fake = [
(torch.zeros(1, 64, 512), torch.LongTensor([1, 2, 3]), 3, "ABC"),
(torch.zeros(1, 64, 512), torch.LongTensor([4, 5]), 2, "DE"),
(torch.zeros(1, 64, 512), torch.LongTensor([6, 7, 8, 9]), 4, "FGHI"),
]
imgs, tgts, tlens, txts = collate_fn(fake)
print(f"\n collate_fn : images={imgs.shape} "
f"targets={tgts.shape} lengths={tlens.tolist()}")
print("\n β All checks passed.\n") |