Spaces:
Sleeping
Sleeping
File size: 19,067 Bytes
d7f1cd8 | 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 | """
Document Readability Scorer
============================
A multi-signal pre-screening system for document validation pipelines.
Scores documents on readability before expensive OCR/LLM inference.
Signals extracted (all normalized to 0-1, higher = better):
1. Sharpness β Laplacian variance + FFT high-freq energy
2. Contrast β RMS contrast + Michelson contrast
3. Noise level β Estimated noise sigma (inverted: low noise = high score)
4. Text presence β MSER-based text region coverage + edge density
5. Brightness β Penalizes over/under-exposed documents
6. Entropy β Shannon entropy (blank pages score low)
7. Learned IQA β CLIP-IQA or BRISQUE via pyiqa (optional, GPU-free)
The composite "readability_score" is a weighted sum of these signals.
Weights are fully configurable for calibration to your pipeline.
Usage:
scorer = DocumentReadabilityScorer()
result = scorer.score("document.png")
print(result["readability_score"]) # float in [0, 1]
print(result["ocr_recommended"]) # bool
print(result["signals"]) # dict of all sub-scores
"""
from __future__ import annotations
import warnings
from dataclasses import dataclass, field
from pathlib import Path
from typing import Optional, Union
import cv2
import numpy as np
from PIL import Image
from scipy import ndimage
from skimage.filters import sobel
from skimage.measure import shannon_entropy
warnings.filterwarnings("ignore", category=UserWarning)
# βββ Configuration βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
@dataclass
class ScorerConfig:
"""Weights and thresholds for the readability scorer.
All weights should sum to 1.0. Adjust these to calibrate
the scorer for your specific document types.
"""
# Signal weights (must sum to 1.0)
w_sharpness: float = 0.30
w_contrast: float = 0.15
w_noise: float = 0.10
w_text_presence: float = 0.15
w_brightness: float = 0.05
w_entropy: float = 0.10
w_learned_iqa: float = 0.15
# Decision threshold
ocr_threshold: float = 0.45 # below this β skip OCR
# Normalization constants (tune per your doc distribution)
laplacian_cap: float = 800.0 # laplacian var at which sharpness = 1.0
noise_cap: float = 15.0 # noise sigma at which noise_score = 0.0
min_text_coverage: float = 0.01 # below this β likely blank
# Learned metric to use (set to None to disable)
learned_metric: Optional[str] = "clipiqa" # "clipiqa", "brisque", "niqe", "topiq_nr", None
# Whether to use GPU for learned metrics
device: str = "cpu"
def validate(self):
total = (self.w_sharpness + self.w_contrast + self.w_noise +
self.w_text_presence + self.w_brightness + self.w_entropy +
self.w_learned_iqa)
if abs(total - 1.0) > 0.01:
raise ValueError(f"Weights must sum to 1.0, got {total:.3f}")
# βββ Signal Extractors ββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _load_gray(image: Union[str, Path, np.ndarray, Image.Image]) -> np.ndarray:
"""Load image as grayscale numpy array."""
if isinstance(image, (str, Path)):
img = cv2.imread(str(image))
if img is None:
raise FileNotFoundError(f"Cannot read image: {image}")
return cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
elif isinstance(image, Image.Image):
return np.array(image.convert("L"))
elif isinstance(image, np.ndarray):
if image.ndim == 3:
return cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
return image
raise TypeError(f"Unsupported image type: {type(image)}")
def _load_color(image: Union[str, Path, np.ndarray, Image.Image]) -> np.ndarray:
"""Load image as BGR numpy array."""
if isinstance(image, (str, Path)):
img = cv2.imread(str(image))
if img is None:
raise FileNotFoundError(f"Cannot read image: {image}")
return img
elif isinstance(image, Image.Image):
return cv2.cvtColor(np.array(image.convert("RGB")), cv2.COLOR_RGB2BGR)
elif isinstance(image, np.ndarray):
return image
raise TypeError(f"Unsupported image type: {type(image)}")
def sharpness_score(gray: np.ndarray, laplacian_cap: float = 800.0) -> dict:
"""
Sharpness via Laplacian variance + FFT high-frequency energy.
Laplacian variance: measures second-derivative magnitude.
- Sharp document text: 200-2000+
- Moderately blurry: 50-200
- Very blurry: <50
FFT energy ratio: fraction of spectral energy in high frequencies.
"""
# Laplacian variance
lap = cv2.Laplacian(gray, cv2.CV_64F)
lap_var = float(lap.var())
lap_norm = min(lap_var / laplacian_cap, 1.0)
# FFT-based: ratio of high-freq energy to total energy
h, w = gray.shape
f = np.fft.fft2(gray.astype(np.float64))
fshift = np.fft.fftshift(f)
magnitude = np.abs(fshift)
total_energy = magnitude.sum()
# Create low-freq mask (center circle, radius = 5% of min dimension)
cy, cx = h // 2, w // 2
radius = int(min(h, w) * 0.05)
Y, X = np.ogrid[:h, :w]
low_freq_mask = ((Y - cy) ** 2 + (X - cx) ** 2) <= radius ** 2
low_energy = magnitude[low_freq_mask].sum()
high_freq_ratio = float(1.0 - low_energy / (total_energy + 1e-10))
# Combined sharpness: 70% Laplacian + 30% FFT
combined = 0.7 * lap_norm + 0.3 * high_freq_ratio
return {
"sharpness": float(np.clip(combined, 0, 1)),
"laplacian_variance": lap_var,
"high_freq_ratio": high_freq_ratio,
}
def contrast_score(gray: np.ndarray) -> dict:
"""
Contrast via RMS and Michelson metrics.
Good documents have RMS contrast ~0.2-0.5 (black text on white).
Washed-out or very dark scans have low contrast.
"""
# RMS contrast
rms = float(gray.std() / 255.0)
# Michelson contrast
i_max, i_min = float(gray.max()), float(gray.min())
michelson = (i_max - i_min) / (i_max + i_min + 1e-10)
# Normalize: RMS of 0.25+ is good for documents
rms_norm = min(rms / 0.30, 1.0)
mich_norm = michelson # already in [0, 1]
combined = 0.6 * rms_norm + 0.4 * mich_norm
return {
"contrast": float(np.clip(combined, 0, 1)),
"rms_contrast": rms,
"michelson_contrast": float(michelson),
}
def noise_score(gray: np.ndarray, noise_cap: float = 15.0) -> dict:
"""
Noise estimation via Immerkær (1996) method.
Uses a 3x3 Laplacian kernel on the image to isolate high-frequency noise.
Clean documents: sigma < 3
Noisy scans: sigma 5-15
Very noisy: sigma > 15
"""
H = np.array([[1, -2, 1], [-2, 4, -2], [1, -2, 1]], dtype=np.float64)
filtered = ndimage.convolve(gray.astype(np.float64), H)
sigma = float(np.abs(filtered).mean() * np.sqrt(np.pi / 2) / 6.0)
# Invert: low noise = high score
noise_norm = 1.0 - min(sigma / noise_cap, 1.0)
return {
"noise": float(np.clip(noise_norm, 0, 1)),
"noise_sigma": sigma,
}
def text_presence_score(gray: np.ndarray, min_coverage: float = 0.01) -> dict:
"""
Text presence via MSER regions + edge density.
MSER (Maximally Stable Extremal Regions) detects text-like blobs.
Edge density via Sobel measures structural content.
"""
# MSER text region detection
mser = cv2.MSER_create()
mser.setDelta(5)
mser.setMinArea(30)
mser.setMaxArea(int(gray.size * 0.05))
mser.setMaxVariation(0.25)
try:
regions, _ = mser.detectRegions(gray)
except cv2.error:
regions = []
if regions:
mask = np.zeros_like(gray)
for r in regions:
hull = cv2.convexHull(r.reshape(-1, 1, 2))
cv2.fillPoly(mask, [hull], 255)
text_coverage = float(mask.sum() / (255.0 * mask.size))
else:
text_coverage = 0.0
# Edge density via Sobel
gray_float = gray.astype(np.float64) / 255.0
edges = sobel(gray_float)
edge_density = float(edges.mean())
# Normalize: coverage >5% is good, edges >0.05 is good
cov_norm = min(text_coverage / 0.10, 1.0)
edge_norm = min(edge_density / 0.08, 1.0)
combined = 0.5 * cov_norm + 0.5 * edge_norm
has_text = text_coverage > min_coverage or edge_density > 0.02
return {
"text_presence": float(np.clip(combined, 0, 1)),
"text_coverage": text_coverage,
"edge_density": edge_density,
"has_text": has_text,
}
def brightness_score(gray: np.ndarray) -> dict:
"""
Brightness assessment β penalizes over/under-exposure.
Ideal document: mean brightness ~160-245 (white paper, dark text).
Score drops for very dark (<80) or fully saturated (==255 everywhere).
Note: Documents naturally have many white pixels (paper background).
White paper with mean brightness ~240-250 is normal and good.
"""
mean_brightness = float(gray.mean())
# Fraction of truly problematic pixels
dark_frac = float((gray < 15).sum() / gray.size) # crushed to black
pure_white_frac = float((gray == 255).sum() / gray.size) # fully saturated
# Score mapping for documents:
# Very dark (<60): bad
# Dim (60-140): mediocre
# Normal (140-250): good (peak at 200-220, but 240-250 is still fine)
# Pure white (>252): suspicious
if mean_brightness < 60:
bright_norm = mean_brightness / 60.0 * 0.3
elif mean_brightness < 140:
bright_norm = 0.3 + (mean_brightness - 60) / 80.0 * 0.5
elif mean_brightness <= 250:
# Wide sweet spot for documents: 140-250 is all good
# Peak at 200, but gentle falloff
dist_from_ideal = abs(mean_brightness - 200) / 60.0
bright_norm = 1.0 - dist_from_ideal * 0.2 # at 250: 0.83, at 140: 0.80
else:
# Over 250 β nearly blank white
bright_norm = max(0.4, 1.0 - (mean_brightness - 250) / 5.0)
# Only penalize if image is mostly crushed blacks or ALL pure white
# (pure_white_frac of 0.9 on a text doc is fine β paper is white)
exposure_penalty = min(dark_frac * 3 + max(0, pure_white_frac - 0.95) * 5, 0.5)
bright_norm = max(0, bright_norm - exposure_penalty)
return {
"brightness": float(np.clip(bright_norm, 0, 1)),
"mean_brightness": mean_brightness,
"dark_pixel_frac": dark_frac,
"bright_pixel_frac": pure_white_frac,
}
def entropy_score(gray: np.ndarray) -> dict:
"""
Shannon entropy β measures information content.
Blank/uniform pages: entropy ~0-3
Text documents: entropy ~5-7
Complex images: entropy ~7-8
"""
ent = float(shannon_entropy(gray))
# Normalize: entropy of 4+ is good for documents (lower threshold than natural images)
# Blank page: ~0-2, simple doc: 3-5, rich doc: 5-7
ent_norm = min(ent / 5.5, 1.0)
return {
"entropy": float(np.clip(ent_norm, 0, 1)),
"shannon_entropy": ent,
}
# βββ Learned IQA (optional) βββββββββββββββββββββββββββββββββββββββββββββββββ
_iqa_cache: dict = {}
def learned_iqa_score(
image: Union[str, Path, np.ndarray, Image.Image],
metric_name: str = "clipiqa",
device: str = "cpu",
) -> dict:
"""
Learned no-reference IQA via pyiqa library.
Supported metrics (all run on CPU):
- clipiqa: CLIP-IQA (0-1, higher=better)
- brisque: BRISQUE (0-100, lower=better, we invert)
- niqe: NIQE (lower=better, we invert)
- topiq_nr: TOPIQ-NR (0-1, higher=better)
"""
import torch
import pyiqa
cache_key = f"{metric_name}_{device}"
if cache_key not in _iqa_cache:
_iqa_cache[cache_key] = pyiqa.create_metric(metric_name, device=device)
metric = _iqa_cache[cache_key]
lower_better = metric.lower_better
# Convert to tensor
if isinstance(image, (str, Path)):
pil_img = Image.open(str(image)).convert("RGB")
elif isinstance(image, np.ndarray):
if image.ndim == 2:
pil_img = Image.fromarray(image).convert("RGB")
else:
pil_img = Image.fromarray(cv2.cvtColor(image, cv2.COLOR_BGR2RGB))
elif isinstance(image, Image.Image):
pil_img = image.convert("RGB")
else:
raise TypeError(f"Unsupported type: {type(image)}")
# Resize for speed (IQA doesn't need full resolution)
max_dim = 512
w, h = pil_img.size
if max(w, h) > max_dim:
scale = max_dim / max(w, h)
pil_img = pil_img.resize((int(w * scale), int(h * scale)), Image.LANCZOS)
img_tensor = torch.from_numpy(
np.array(pil_img).transpose(2, 0, 1)
).float().unsqueeze(0) / 255.0
img_tensor = img_tensor.to(device)
with torch.no_grad():
raw_score = float(metric(img_tensor).item())
# Normalize to [0, 1] higher=better
if lower_better:
if metric_name == "brisque":
normalized = float(np.clip(1.0 - raw_score / 100.0, 0, 1))
elif metric_name == "niqe":
normalized = float(np.clip(1.0 - raw_score / 20.0, 0, 1))
else:
normalized = float(np.clip(1.0 - raw_score / 50.0, 0, 1))
else:
normalized = float(np.clip(raw_score, 0, 1))
return {
"learned_iqa": normalized,
f"{metric_name}_raw": raw_score,
"metric_name": metric_name,
}
# βββ Main Scorer βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
@dataclass
class ReadabilityResult:
"""Complete readability assessment for a document image."""
readability_score: float # Composite score [0, 1]
ocr_recommended: bool # Whether to proceed with OCR
confidence_label: str # "excellent" / "good" / "fair" / "poor" / "bad"
signals: dict # All individual signal scores and raw values
config: dict # Config used for this scoring
def to_dict(self) -> dict:
return {
"readability_score": self.readability_score,
"ocr_recommended": self.ocr_recommended,
"confidence_label": self.confidence_label,
"signals": self.signals,
}
class DocumentReadabilityScorer:
"""
Multi-signal document readability scorer.
Example:
scorer = DocumentReadabilityScorer()
result = scorer.score("scan.pdf")
if result.ocr_recommended:
run_ocr(...)
else:
log_rejected(result.signals)
"""
def __init__(self, config: Optional[ScorerConfig] = None):
self.config = config or ScorerConfig()
self.config.validate()
def score(
self,
image: Union[str, Path, np.ndarray, Image.Image],
) -> ReadabilityResult:
"""
Score a document image for readability.
Args:
image: File path, numpy array (BGR or gray), or PIL Image.
Returns:
ReadabilityResult with composite score, sub-signals, and recommendation.
"""
cfg = self.config
gray = _load_gray(image)
# Extract all classical signals
sharp = sharpness_score(gray, cfg.laplacian_cap)
cont = contrast_score(gray)
noi = noise_score(gray, cfg.noise_cap)
text = text_presence_score(gray, cfg.min_text_coverage)
bright = brightness_score(gray)
ent = entropy_score(gray)
# Optional learned IQA
if cfg.learned_metric:
try:
iqa = learned_iqa_score(image, cfg.learned_metric, cfg.device)
except Exception as e:
# Fall back gracefully β redistribute weight to sharpness
iqa = {"learned_iqa": 0.5, "error": str(e), "metric_name": cfg.learned_metric}
else:
iqa = {"learned_iqa": 0.5, "metric_name": "disabled"}
# Composite score
composite = (
cfg.w_sharpness * sharp["sharpness"] +
cfg.w_contrast * cont["contrast"] +
cfg.w_noise * noi["noise"] +
cfg.w_text_presence * text["text_presence"] +
cfg.w_brightness * bright["brightness"] +
cfg.w_entropy * ent["entropy"] +
cfg.w_learned_iqa * iqa["learned_iqa"]
)
composite = float(np.clip(composite, 0, 1))
# Label
if composite >= 0.80:
label = "excellent"
elif composite >= 0.60:
label = "good"
elif composite >= 0.40:
label = "fair"
elif composite >= 0.20:
label = "poor"
else:
label = "bad"
# Merge all signals
signals = {}
for d in [sharp, cont, noi, text, bright, ent, iqa]:
signals.update(d)
return ReadabilityResult(
readability_score=round(composite, 4),
ocr_recommended=composite >= cfg.ocr_threshold,
confidence_label=label,
signals=signals,
config={
"weights": {
"sharpness": cfg.w_sharpness,
"contrast": cfg.w_contrast,
"noise": cfg.w_noise,
"text_presence": cfg.w_text_presence,
"brightness": cfg.w_brightness,
"entropy": cfg.w_entropy,
"learned_iqa": cfg.w_learned_iqa,
},
"ocr_threshold": cfg.ocr_threshold,
"learned_metric": cfg.learned_metric or "disabled",
},
)
# βββ Batch processing helper βββββββββββββββββββββββββββββββββββββββββββββββββ
def score_batch(
image_paths: list[Union[str, Path]],
config: Optional[ScorerConfig] = None,
sort_by_score: bool = True,
) -> list[dict]:
"""Score a batch of documents and optionally sort by readability."""
scorer = DocumentReadabilityScorer(config)
results = []
for path in image_paths:
try:
result = scorer.score(path)
results.append({
"path": str(path),
**result.to_dict(),
})
except Exception as e:
results.append({
"path": str(path),
"readability_score": 0.0,
"ocr_recommended": False,
"confidence_label": "error",
"error": str(e),
})
if sort_by_score:
results.sort(key=lambda x: x["readability_score"], reverse=True)
return results |