Spaces:
Running
Running
File size: 32,538 Bytes
7931476 d895a0c 5be749f d895a0c 433e26f d895a0c 5be749f d895a0c 433e26f d895a0c 7931476 d895a0c 433e26f d895a0c 7931476 d895a0c 433e26f d895a0c 433e26f d895a0c 7931476 d895a0c 7931476 d895a0c 433e26f d895a0c 433e26f d895a0c 433e26f d895a0c 7931476 d895a0c 433e26f d895a0c 7931476 d895a0c 7931476 d895a0c 7931476 d895a0c 7931476 d895a0c 7931476 433e26f d895a0c 7931476 d895a0c 433e26f d895a0c 7931476 d895a0c 7931476 d895a0c 7931476 d895a0c 433e26f d895a0c 7931476 d895a0c 7931476 d895a0c 7931476 d895a0c 433e26f 7931476 d895a0c 433e26f d895a0c 7931476 d895a0c 433e26f d895a0c 433e26f d895a0c 7931476 d895a0c 7931476 d895a0c 433e26f d895a0c 433e26f d895a0c 433e26f d895a0c 7931476 d895a0c 433e26f d895a0c 7931476 433e26f 7931476 d895a0c 433e26f d895a0c 433e26f d895a0c 7931476 d895a0c 5be749f d895a0c 5be749f d895a0c 7931476 5be749f 7931476 d895a0c 7931476 d895a0c 433e26f d895a0c 433e26f d895a0c 5be749f d895a0c 5be749f 7931476 d895a0c 7931476 d895a0c 433e26f d895a0c 5be749f d895a0c 7931476 d895a0c 433e26f d895a0c 5be749f d895a0c 5be749f d895a0c | 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 | """Neural face verification, distortion detection, and restoration pipeline.
End-to-end system that:
1. Detects face distortions (blur, beauty filters, compression, warping, etc.)
2. Classifies distortion type and severity using no-reference quality metrics
3. Restores faces using cascaded neural networks (CodeFormer β GFPGAN β Real-ESRGAN)
4. Verifies output identity matches input via ArcFace embeddings
5. Scores output realism using learned perceptual metrics
Designed for:
- Cleaning scraped training data (reject/fix bad images before pair generation)
- Post-diffusion quality gate (ensure generated faces pass realism threshold)
- Filter removal (undo Snapchat/Instagram beauty filters for clinical use)
"""
from __future__ import annotations
import logging
from dataclasses import dataclass, field
from pathlib import Path
from typing import Any
import cv2
import numpy as np
logger = logging.getLogger(__name__)
# ---------------------------------------------------------------------------
# Data structures
# ---------------------------------------------------------------------------
@dataclass
class DistortionReport:
"""Analysis of detected distortions in a face image."""
# Overall quality score (0-100, higher = better)
quality_score: float = 0.0
# Individual distortion scores (0-1, higher = more distorted)
blur_score: float = 0.0 # Laplacian variance-based
noise_score: float = 0.0 # High-freq energy ratio
compression_score: float = 0.0 # JPEG block artifact detection
oversmooth_score: float = 0.0 # Beauty filter / airbrushed detection
color_cast_score: float = 0.0 # Unnatural color shift
geometric_distort: float = 0.0 # Face proportion anomalies
lighting_score: float = 0.0 # Over/under exposure
# Classification
primary_distortion: str = "none"
severity: str = "none" # none, mild, moderate, severe
is_usable: bool = True # Whether image is worth restoring vs rejecting
# Details
details: dict = field(default_factory=dict)
def summary(self) -> str:
lines = [
f"Quality Score: {self.quality_score:.1f}/100",
f"Primary Issue: {self.primary_distortion} ({self.severity})",
f"Usable: {self.is_usable}",
"",
"Distortion Breakdown:",
f" Blur: {self.blur_score:.3f}",
f" Noise: {self.noise_score:.3f}",
f" Compression: {self.compression_score:.3f}",
f" Oversmooth: {self.oversmooth_score:.3f}",
f" Color Cast: {self.color_cast_score:.3f}",
f" Geometric: {self.geometric_distort:.3f}",
f" Lighting: {self.lighting_score:.3f}",
]
return "\n".join(lines)
@dataclass
class RestorationResult:
"""Result of neural face restoration pipeline."""
restored: np.ndarray # Restored BGR image
original: np.ndarray # Original BGR image
distortion_report: DistortionReport # Pre-restoration analysis
post_quality_score: float = 0.0 # Quality after restoration
identity_similarity: float = 0.0 # ArcFace cosine sim (original vs restored)
identity_preserved: bool = True # Whether identity check passed
restoration_stages: list[str] = field(default_factory=list) # Which nets ran
improvement: float = 0.0 # quality_after - quality_before
def summary(self) -> str:
lines = [
f"Pre-restoration: {self.distortion_report.quality_score:.1f}/100",
f"Post-restoration: {self.post_quality_score:.1f}/100",
f"Improvement: +{self.improvement:.1f}",
f"Identity Sim: {self.identity_similarity:.3f}",
f"Identity OK: {self.identity_preserved}",
f"Stages Used: {' β '.join(self.restoration_stages) or 'none'}",
]
return "\n".join(lines)
@dataclass
class BatchVerificationReport:
"""Summary of batch face verification/restoration."""
total: int = 0
passed: int = 0 # Good quality, no fix needed
restored: int = 0 # Fixed and now usable
rejected: int = 0 # Too distorted to salvage
identity_failures: int = 0 # Restoration changed identity
avg_quality_before: float = 0.0
avg_quality_after: float = 0.0
avg_identity_sim: float = 0.0
distortion_counts: dict[str, int] = field(default_factory=dict)
def summary(self) -> str:
lines = [
f"Total Images: {self.total}",
f" Passed (good): {self.passed}",
f" Restored: {self.restored}",
f" Rejected: {self.rejected}",
f" Identity Fail: {self.identity_failures}",
f"Avg Quality Before: {self.avg_quality_before:.1f}",
f"Avg Quality After: {self.avg_quality_after:.1f}",
f"Avg Identity Sim: {self.avg_identity_sim:.3f}",
"",
"Distortion Breakdown:",
]
for dist_type, count in sorted(
self.distortion_counts.items(),
key=lambda x: -x[1],
):
lines.append(f" {dist_type}: {count}")
return "\n".join(lines)
# ---------------------------------------------------------------------------
# Distortion Detection (classical + neural)
# ---------------------------------------------------------------------------
def detect_blur(image: np.ndarray) -> float:
"""Detect blur using Laplacian variance.
Low variance = blurry. We normalize to 0-1 where 1 = very blurry.
Uses both Laplacian variance and gradient magnitude for robustness.
"""
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY) if image.ndim == 3 else image
# Laplacian variance (primary metric)
lap_var = cv2.Laplacian(gray, cv2.CV_64F).var()
# Gradient magnitude (secondary)
gx = cv2.Sobel(gray, cv2.CV_64F, 1, 0, ksize=3)
gy = cv2.Sobel(gray, cv2.CV_64F, 0, 1, ksize=3)
grad_mag = np.sqrt(gx**2 + gy**2).mean()
# Normalize: typical sharp face has lap_var > 500, grad_mag > 30
blur_lap = 1.0 - min(lap_var / 800.0, 1.0)
blur_grad = 1.0 - min(grad_mag / 50.0, 1.0)
return float(np.clip(0.6 * blur_lap + 0.4 * blur_grad, 0, 1))
def detect_noise(image: np.ndarray) -> float:
"""Detect image noise level.
Estimates noise by measuring high-frequency energy in smooth regions.
Uses the median absolute deviation of the Laplacian (robust estimator).
"""
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY) if image.ndim == 3 else image
# Robust noise estimation via MAD of Laplacian
lap = cv2.Laplacian(gray.astype(np.float64), cv2.CV_64F)
sigma_est = np.median(np.abs(lap)) * 1.4826 # MAD β std conversion
# Normalize: sigma > 20 is very noisy
return float(np.clip(sigma_est / 25.0, 0, 1))
def detect_compression_artifacts(image: np.ndarray) -> float:
"""Detect JPEG compression block artifacts.
Measures energy at 8x8 block boundaries (JPEG DCT block size).
High boundary energy relative to interior = compression artifacts.
"""
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY) if image.ndim == 3 else image
h, w = gray.shape
if h < 16 or w < 16:
return 0.0
gray_f = gray.astype(np.float64)
# Compute horizontal and vertical differences
h_diff = np.abs(np.diff(gray_f, axis=1))
v_diff = np.abs(np.diff(gray_f, axis=0))
# Energy at 8-pixel boundaries vs non-boundaries
h_boundary = h_diff[:, 7::8].mean() if h_diff[:, 7::8].size > 0 else 0
h_interior = h_diff.mean()
v_boundary = v_diff[7::8, :].mean() if v_diff[7::8, :].size > 0 else 0
v_interior = v_diff.mean()
if h_interior < 1e-6 or v_interior < 1e-6:
return 0.0
# Ratio of boundary to interior energy (>1 means block artifacts)
h_ratio = h_boundary / (h_interior + 1e-6)
v_ratio = v_boundary / (v_interior + 1e-6)
artifact_ratio = (h_ratio + v_ratio) / 2.0
# Normalize: ratio > 1.5 indicates visible artifacts
return float(np.clip((artifact_ratio - 1.0) / 0.8, 0, 1))
def detect_oversmoothing(image: np.ndarray) -> float:
"""Detect beauty filter / airbrushed skin (oversmoothing).
Beauty filters remove skin texture while preserving edges. We detect
this by measuring the ratio of edge energy to texture energy.
High edge / low texture = beauty filtered.
"""
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY) if image.ndim == 3 else image
h, w = gray.shape
# Focus on face center region (avoid background)
if h < 8 or w < 8:
return 0.0 # Too small to analyze
roi = gray[h // 4 : 3 * h // 4, w // 4 : 3 * w // 4]
# Texture energy: variance of high-pass filtered image
blurred = cv2.GaussianBlur(roi.astype(np.float64), (0, 0), 2.0)
high_pass = roi.astype(np.float64) - blurred
texture_energy = np.var(high_pass)
# Edge energy: Canny edge density
edges = cv2.Canny(roi, 50, 150)
edge_density = np.mean(edges > 0)
# Oversmooth: low texture but edges still present
# Natural skin: texture_energy > 20, beauty filter: < 8
smooth_score = 1.0 - min(texture_energy / 30.0, 1.0)
# If there are still strong edges but no texture, it's a filter
if edge_density > 0.02:
smooth_score *= 1.3 # Amplify if edges present but no texture
return float(np.clip(smooth_score, 0, 1))
def detect_color_cast(image: np.ndarray) -> float:
"""Detect unnatural color cast (Instagram-style filters).
Measures deviation of average A/B channels in LAB space from
neutral. Natural skin has consistent LAB distributions; filtered
images shift these channels.
"""
lab = cv2.cvtColor(image, cv2.COLOR_BGR2LAB).astype(np.float32)
h, w = image.shape[:2]
# Sample face center region
roi = lab[h // 4 : 3 * h // 4, w // 4 : 3 * w // 4]
# A channel: green-red axis (neutral ~128)
# B channel: blue-yellow axis (neutral ~128)
a_mean = roi[:, :, 1].mean()
b_mean = roi[:, :, 2].mean()
# Deviation from neutral
a_dev = abs(a_mean - 128) / 128.0
b_dev = abs(b_mean - 128) / 128.0
# Also check if color distribution is unnaturally narrow (saturated filter)
a_std = roi[:, :, 1].std()
b_std = roi[:, :, 2].std()
narrow_color = max(0, 1.0 - (a_std + b_std) / 30.0)
score = 0.5 * (a_dev + b_dev) + 0.3 * narrow_color
return float(np.clip(score, 0, 1))
def detect_geometric_distortion(image: np.ndarray) -> float:
"""Detect geometric face distortion (warping filters, lens distortion).
Uses MediaPipe landmarks to check face proportions against anatomical
norms. Distorted faces have abnormal inter-ocular / face-width ratios.
"""
try:
from landmarkdiff.landmarks import extract_landmarks
except ImportError:
return 0.0
face = extract_landmarks(image)
if face is None:
return 0.5 # Can't detect face = possibly distorted
coords = face.pixel_coords
h, w = image.shape[:2]
if len(coords) < 478:
return 0.5 # Incomplete landmark set
# Key ratios that should be anatomically consistent
left_eye = coords[33]
right_eye = coords[263]
nose_tip = coords[1]
chin = coords[152]
forehead = coords[10]
iod = np.linalg.norm(left_eye - right_eye)
face_height = np.linalg.norm(forehead - chin)
nose_to_chin = np.linalg.norm(nose_tip - chin)
if iod < 1.0 or face_height < 1.0:
return 0.5
# Anatomical norms (approximate):
# face_height / iod β 2.5-3.5
# nose_to_chin / face_height β 0.3-0.45
height_ratio = face_height / iod
lower_ratio = nose_to_chin / face_height
# Score deviations from normal ranges
height_dev = max(0, abs(height_ratio - 3.0) - 0.5) / 1.5
lower_dev = max(0, abs(lower_ratio - 0.38) - 0.08) / 0.15
# Eye symmetry check (vertical alignment)
eye_tilt = abs(left_eye[1] - right_eye[1]) / (iod + 1e-6)
tilt_dev = max(0, eye_tilt - 0.05) / 0.15
score = 0.4 * height_dev + 0.3 * lower_dev + 0.3 * tilt_dev
return float(np.clip(score, 0, 1))
def detect_lighting_issues(image: np.ndarray) -> float:
"""Detect over/under exposure and harsh lighting.
Checks luminance histogram for clipping and uneven distribution.
"""
lab = cv2.cvtColor(image, cv2.COLOR_BGR2LAB)
l_channel = lab[:, :, 0]
# Check for clipping
overexposed = np.mean(l_channel > 245) * 5 # Fraction near white
underexposed = np.mean(l_channel < 10) * 5 # Fraction near black
# Check for bimodal distribution (harsh shadows)
hist = cv2.calcHist([l_channel], [0], None, [256], [0, 256]).flatten()
hist_sum = hist.sum()
if hist_sum < 1e-10:
return 0.0
hist = hist / hist_sum
# Measure how spread out the histogram is
entropy = -np.sum(hist[hist > 0] * np.log2(hist[hist > 0] + 1e-10))
# Low entropy = concentrated = potentially problematic
entropy_score = max(0, 1.0 - entropy / 7.0)
score = 0.4 * overexposed + 0.4 * underexposed + 0.2 * entropy_score
return float(np.clip(score, 0, 1))
def analyze_distortions(image: np.ndarray) -> DistortionReport:
"""Run full distortion analysis on a face image.
Combines all detection methods into a comprehensive report with
quality score, primary distortion classification, and severity.
"""
blur = detect_blur(image)
noise = detect_noise(image)
compression = detect_compression_artifacts(image)
oversmooth = detect_oversmoothing(image)
color_cast = detect_color_cast(image)
geometric = detect_geometric_distortion(image)
lighting = detect_lighting_issues(image)
# Overall quality: weighted combination (inverted β 100 = perfect)
weighted = (
0.25 * blur
+ 0.15 * noise
+ 0.10 * compression
+ 0.20 * oversmooth
+ 0.10 * color_cast
+ 0.10 * geometric
+ 0.10 * lighting
)
quality = (1.0 - weighted) * 100.0
# Classify primary distortion
scores = {
"blur": blur,
"noise": noise,
"compression": compression,
"oversmooth": oversmooth,
"color_cast": color_cast,
"geometric": geometric,
"lighting": lighting,
}
primary = max(scores, key=scores.get)
primary_val = scores[primary]
if primary_val < 0.15:
severity = "none"
primary = "none"
elif primary_val < 0.35:
severity = "mild"
elif primary_val < 0.60:
severity = "moderate"
else:
severity = "severe"
# Image is usable if quality > 30 and no severe geometric distortion
is_usable = quality > 25 and geometric < 0.7
return DistortionReport(
quality_score=quality,
blur_score=blur,
noise_score=noise,
compression_score=compression,
oversmooth_score=oversmooth,
color_cast_score=color_cast,
geometric_distort=geometric,
lighting_score=lighting,
primary_distortion=primary,
severity=severity,
is_usable=is_usable,
details=scores,
)
# ---------------------------------------------------------------------------
# Neural Face Quality Scoring (no-reference)
# ---------------------------------------------------------------------------
_FACE_QUALITY_NET = None
def _get_face_quality_scorer() -> Any:
"""Get or create singleton face quality assessment model.
Uses FaceXLib's quality scorer or falls back to BRISQUE-style features.
"""
global _FACE_QUALITY_NET
if _FACE_QUALITY_NET is not None:
return _FACE_QUALITY_NET
try:
from facexlib.assessment import init_assessment_model
_FACE_QUALITY_NET = init_assessment_model("hypernet")
return _FACE_QUALITY_NET
except Exception:
pass
return None
def neural_quality_score(image: np.ndarray) -> float:
"""Score face quality using neural network (0-100, higher = better).
Tries FaceXLib quality assessment first, then falls back to
BRISQUE-style scoring using OpenCV's QualityBRISQUE if available,
or classical metrics as last resort.
"""
# Try neural scorer
scorer = _get_face_quality_scorer()
if scorer is not None:
try:
import torch
from facexlib.utils import img2tensor
img_t = img2tensor(image / 255.0, bgr2rgb=True, float32=True)
img_t = img_t.unsqueeze(0)
if torch.cuda.is_available():
img_t = img_t.cuda()
scorer = scorer.cuda()
with torch.no_grad():
score = scorer(img_t).item()
return float(np.clip(score * 100, 0, 100))
except Exception:
pass
# Fallback: composite classical score
report = analyze_distortions(image)
return report.quality_score
# ---------------------------------------------------------------------------
# Neural Face Restoration (cascaded)
# ---------------------------------------------------------------------------
def restore_face(
image: np.ndarray,
distortion: DistortionReport | None = None,
mode: str = "auto",
codeformer_fidelity: float = 0.7,
) -> tuple[np.ndarray, list[str]]:
"""Cascaded neural face restoration.
Selects and applies restoration networks based on detected distortions:
- Blur/oversmooth β CodeFormer (recovers texture from codebook)
- Noise/compression β GFPGAN (trained on degraded faces)
- Background β Real-ESRGAN (neural 4x upscale + downsample)
- Color cast β Classical LAB correction (no neural net needed)
- Geometric β Not fixable by restoration (flag and skip)
Args:
image: BGR face image to restore.
distortion: Pre-computed distortion report (computed if None).
mode: 'auto' (choose based on distortion), 'codeformer', 'gfpgan', 'all'.
codeformer_fidelity: CodeFormer quality-fidelity tradeoff.
Returns:
Tuple of (restored BGR image, list of stages applied).
"""
if distortion is None:
distortion = analyze_distortions(image)
result = image.copy()
stages = []
# Step 0: Fix color cast first (classical β fast, doesn't affect identity)
if distortion.color_cast_score > 0.25:
result = _fix_color_cast(result)
stages.append("color_correction")
# Step 1: Fix lighting issues (classical)
if distortion.lighting_score > 0.35:
result = _fix_lighting(result)
stages.append("lighting_fix")
# Step 2: Neural face restoration
if mode == "auto":
# Choose based on what's wrong
needs_face_restore = (
distortion.blur_score > 0.2
or distortion.oversmooth_score > 0.25
or distortion.noise_score > 0.25
or distortion.compression_score > 0.2
)
if needs_face_restore:
mode = "codeformer" # CodeFormer handles most degradations well
if mode in ("codeformer", "all"):
restored = _try_codeformer(result, fidelity=codeformer_fidelity)
if restored is not None:
result = restored
stages.append("codeformer")
else:
# Fallback to GFPGAN
restored = _try_gfpgan(result)
if restored is not None:
result = restored
stages.append("gfpgan")
elif mode == "gfpgan":
restored = _try_gfpgan(result)
if restored is not None:
result = restored
stages.append("gfpgan")
# Step 3: Background enhancement with Real-ESRGAN (if image is low-res)
h, w = result.shape[:2]
if h < 400 or w < 400:
enhanced = _try_realesrgan(result)
if enhanced is not None:
result = enhanced
stages.append("realesrgan")
# Step 4: Mild sharpening if still soft after restoration
post_blur = detect_blur(result)
if post_blur > 0.3:
from landmarkdiff.postprocess import frequency_aware_sharpen
result = frequency_aware_sharpen(result, strength=0.3)
stages.append("sharpen")
return result, stages
def _try_codeformer(image: np.ndarray, fidelity: float = 0.7) -> np.ndarray | None:
"""Try CodeFormer restoration. Returns None if unavailable."""
try:
from landmarkdiff.postprocess import restore_face_codeformer
restored = restore_face_codeformer(image, fidelity=fidelity)
if restored is not image:
return restored
except Exception:
pass
return None
def _try_gfpgan(image: np.ndarray) -> np.ndarray | None:
"""Try GFPGAN restoration. Returns None if unavailable."""
try:
from landmarkdiff.postprocess import restore_face_gfpgan
restored = restore_face_gfpgan(image)
if restored is not image:
return restored
except Exception:
pass
return None
_FV_REALESRGAN = None
def _try_realesrgan(image: np.ndarray) -> np.ndarray | None:
"""Try Real-ESRGAN 2x upscale + downsample. Returns None if unavailable."""
try:
import torch
from basicsr.archs.rrdbnet_arch import RRDBNet
from realesrgan import RealESRGANer
global _FV_REALESRGAN
if _FV_REALESRGAN is None:
model = RRDBNet(
num_in_ch=3,
num_out_ch=3,
num_feat=64,
num_block=23,
num_grow_ch=32,
scale=4,
)
_FV_REALESRGAN = RealESRGANer(
scale=4,
model_path="https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth",
model=model,
tile=400,
tile_pad=10,
pre_pad=0,
half=torch.cuda.is_available(),
)
enhanced, _ = _FV_REALESRGAN.enhance(image, outscale=2)
# Downsample to 512x512 for pipeline consistency
enhanced = cv2.resize(enhanced, (512, 512), interpolation=cv2.INTER_LANCZOS4)
return enhanced
except Exception:
pass
return None
def _fix_color_cast(image: np.ndarray) -> np.ndarray:
"""Remove color cast by normalizing A/B channels in LAB space."""
lab = cv2.cvtColor(image, cv2.COLOR_BGR2LAB).astype(np.float32)
# Center A and B channels around 128 (neutral)
for ch in [1, 2]:
channel = lab[:, :, ch]
mean_val = channel.mean()
# Shift toward neutral, but only partially to preserve natural skin tone
shift = (128.0 - mean_val) * 0.6
lab[:, :, ch] = np.clip(channel + shift, 0, 255)
return cv2.cvtColor(lab.astype(np.uint8), cv2.COLOR_LAB2BGR)
def _fix_lighting(image: np.ndarray) -> np.ndarray:
"""Fix over/under exposure using adaptive CLAHE in LAB space."""
lab = cv2.cvtColor(image, cv2.COLOR_BGR2LAB)
# CLAHE on luminance channel only
clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))
lab[:, :, 0] = clahe.apply(lab[:, :, 0])
return cv2.cvtColor(lab, cv2.COLOR_LAB2BGR)
# ---------------------------------------------------------------------------
# ArcFace Identity Verification
# ---------------------------------------------------------------------------
_ARCFACE_APP = None
def _get_arcface() -> Any:
"""Get or create singleton ArcFace model."""
global _ARCFACE_APP
if _ARCFACE_APP is not None:
return _ARCFACE_APP
try:
import torch
from insightface.app import FaceAnalysis
app = FaceAnalysis(
name="buffalo_l",
providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
)
ctx_id = 0 if torch.cuda.is_available() else -1
app.prepare(ctx_id=ctx_id, det_size=(320, 320))
_ARCFACE_APP = app
return app
except Exception:
return None
def get_face_embedding(image: np.ndarray) -> np.ndarray | None:
"""Extract ArcFace 512-d embedding from a face image.
Returns None if no face detected or InsightFace unavailable.
"""
app = _get_arcface()
if app is None:
return None
try:
faces = app.get(image)
if faces:
emb = faces[0].embedding
if np.linalg.norm(emb) < 1e-6:
logger.warning("ArcFace returned near-zero embedding (occluded face?)")
return None
return emb
except Exception:
pass
return None
def verify_identity(
original: np.ndarray,
restored: np.ndarray,
threshold: float = 0.5,
) -> tuple[float, bool]:
"""Compare identity between original and restored using ArcFace.
Returns (cosine_similarity, passed).
Similarity > threshold means same person (0.5 accommodates non-frontal poses).
"""
emb_orig = get_face_embedding(original)
emb_rest = get_face_embedding(restored)
if emb_orig is None or emb_rest is None:
return -1.0, True # Can't verify β assume OK
sim = float(
np.dot(emb_orig, emb_rest) / (np.linalg.norm(emb_orig) * np.linalg.norm(emb_rest) + 1e-8)
)
sim = float(np.clip(sim, -1, 1))
return sim, sim >= threshold
# ---------------------------------------------------------------------------
# Full Verification + Restoration Pipeline
# ---------------------------------------------------------------------------
def verify_and_restore(
image: np.ndarray,
quality_threshold: float = 60.0,
identity_threshold: float = 0.5,
restore_mode: str = "auto",
codeformer_fidelity: float = 0.7,
) -> RestorationResult:
"""Full pipeline: analyze -> restore -> verify identity.
This is the main entry point for the face verifier. It:
1. Analyzes the input for distortions
2. If quality is below threshold, applies neural restoration
3. Verifies the restored face preserves identity
4. Returns comprehensive result with metrics
Args:
image: BGR face image.
quality_threshold: Min quality to skip restoration (0-100).
identity_threshold: Min ArcFace similarity to pass (0-1).
restore_mode: 'auto', 'codeformer', 'gfpgan', 'all'.
codeformer_fidelity: CodeFormer quality-fidelity balance.
Returns:
RestorationResult with restored image and full metrics.
"""
# Step 1: Analyze distortions
report = analyze_distortions(image)
# Step 2: Decide if restoration needed
if report.quality_score >= quality_threshold and report.severity in ("none", "mild"):
# Image is good enough β no restoration needed
return RestorationResult(
restored=image.copy(),
original=image.copy(),
distortion_report=report,
post_quality_score=report.quality_score,
identity_similarity=1.0,
identity_preserved=True,
restoration_stages=[],
improvement=0.0,
)
if not report.is_usable:
# Too distorted to salvage
return RestorationResult(
restored=image.copy(),
original=image.copy(),
distortion_report=report,
post_quality_score=report.quality_score,
identity_similarity=0.0,
identity_preserved=False,
restoration_stages=["rejected"],
improvement=0.0,
)
# Step 3: Neural restoration
restored, stages = restore_face(
image,
distortion=report,
mode=restore_mode,
codeformer_fidelity=codeformer_fidelity,
)
# Step 4: Post-restoration quality check
post_quality = neural_quality_score(restored)
# Step 5: Identity verification
sim, id_ok = verify_identity(image, restored, threshold=identity_threshold)
return RestorationResult(
restored=restored,
original=image.copy(),
distortion_report=report,
post_quality_score=post_quality,
identity_similarity=sim,
identity_preserved=id_ok,
restoration_stages=stages,
improvement=post_quality - report.quality_score,
)
# ---------------------------------------------------------------------------
# Batch Processing
# ---------------------------------------------------------------------------
def verify_batch(
image_dir: str,
output_dir: str | None = None,
quality_threshold: float = 60.0,
identity_threshold: float = 0.5,
restore_mode: str = "auto",
save_rejected: bool = False,
extensions: tuple[str, ...] = (".jpg", ".jpeg", ".png", ".webp", ".bmp"),
) -> BatchVerificationReport:
"""Process a directory of face images: analyze, restore, verify, sort.
Outputs:
- {output_dir}/passed/ β good images (no fix needed)
- {output_dir}/restored/ β fixed images
- {output_dir}/rejected/ β too distorted to use (if save_rejected=True)
- {output_dir}/report.txt β batch verification report
Args:
image_dir: Directory of face images to process.
output_dir: Where to save results (default: {image_dir}_verified/).
quality_threshold: Min quality to pass without restoration.
identity_threshold: Min identity similarity after restoration.
restore_mode: 'auto', 'codeformer', 'gfpgan', 'all'.
save_rejected: Whether to copy rejected images to rejected/ subdir.
extensions: File extensions to process.
Returns:
BatchVerificationReport with summary statistics.
"""
image_path = Path(image_dir)
if output_dir is None:
out_path = image_path.parent / f"{image_path.name}_verified"
else:
out_path = Path(output_dir)
# Create output dirs
passed_dir = out_path / "passed"
restored_dir = out_path / "restored"
rejected_dir = out_path / "rejected"
passed_dir.mkdir(parents=True, exist_ok=True)
restored_dir.mkdir(parents=True, exist_ok=True)
if save_rejected:
rejected_dir.mkdir(parents=True, exist_ok=True)
# Find all images
image_files = sorted(
[f for f in image_path.iterdir() if f.suffix.lower() in extensions and f.is_file()]
)
report = BatchVerificationReport(total=len(image_files))
quality_before = []
quality_after = []
identity_sims = []
for i, img_file in enumerate(image_files):
if (i + 1) % 50 == 0 or i == 0:
logger.info("Processing %d/%d: %s", i + 1, len(image_files), img_file.name)
image = cv2.imread(str(img_file))
if image is None:
report.rejected += 1
continue
# Resize to 512x512 for consistency
image = cv2.resize(image, (512, 512))
# Run verification + restoration
result = verify_and_restore(
image,
quality_threshold=quality_threshold,
identity_threshold=identity_threshold,
restore_mode=restore_mode,
)
quality_before.append(result.distortion_report.quality_score)
quality_after.append(result.post_quality_score)
# Track distortion types
dist_type = result.distortion_report.primary_distortion
report.distortion_counts[dist_type] = report.distortion_counts.get(dist_type, 0) + 1
if not result.distortion_report.is_usable or "rejected" in result.restoration_stages:
report.rejected += 1
if save_rejected:
cv2.imwrite(str(rejected_dir / img_file.name), image)
elif not result.restoration_stages:
# Passed without restoration
report.passed += 1
cv2.imwrite(str(passed_dir / img_file.name), image)
else:
# Restored
if result.identity_preserved:
report.restored += 1
cv2.imwrite(str(restored_dir / img_file.name), result.restored)
identity_sims.append(result.identity_similarity)
else:
report.identity_failures += 1
if save_rejected:
cv2.imwrite(str(rejected_dir / img_file.name), image)
# Compute averages
report.avg_quality_before = float(np.mean(quality_before)) if quality_before else 0.0
report.avg_quality_after = float(np.mean(quality_after)) if quality_after else 0.0
report.avg_identity_sim = float(np.mean(identity_sims)) if identity_sims else 0.0
# Save report
report_text = report.summary()
(out_path / "report.txt").write_text(report_text)
logger.info("\n%s", report_text)
logger.info("Results saved to %s/", out_path)
return report
|