Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,698 +1,253 @@
|
|
| 1 |
"""
|
| 2 |
-
|
| 3 |
-
|
| 4 |
-
This app provides a web interface for detecting and classifying document forgeries.
|
| 5 |
"""
|
| 6 |
|
| 7 |
-
import gradio as gr
|
| 8 |
-
import torch
|
| 9 |
import cv2
|
| 10 |
import numpy as np
|
| 11 |
-
from
|
| 12 |
-
import
|
| 13 |
-
from
|
| 14 |
-
import sys
|
| 15 |
-
from typing import Dict, List, Tuple
|
| 16 |
-
import plotly.graph_objects as go
|
| 17 |
-
|
| 18 |
-
# Add src to path
|
| 19 |
-
sys.path.insert(0, str(Path(__file__).parent))
|
| 20 |
-
|
| 21 |
-
from src.models import get_model
|
| 22 |
-
from src.config import get_config
|
| 23 |
-
from src.data.preprocessing import DocumentPreprocessor
|
| 24 |
-
from src.data.augmentation import DatasetAwareAugmentation
|
| 25 |
-
from src.features.region_extraction import get_mask_refiner, get_region_extractor
|
| 26 |
-
from src.features.feature_extraction import get_feature_extractor
|
| 27 |
-
from src.training.classifier import ForgeryClassifier
|
| 28 |
-
|
| 29 |
-
# Class names
|
| 30 |
-
CLASS_NAMES = {0: 'Copy-Move', 1: 'Splicing', 2: 'Text Substitution'}
|
| 31 |
-
CLASS_COLORS = {
|
| 32 |
-
0: (217, 83, 79), # #d9534f - Muted red
|
| 33 |
-
1: (92, 184, 92), # #5cb85c - Muted green
|
| 34 |
-
2: (65, 105, 225) # #4169E1 - Royal blue
|
| 35 |
-
}
|
| 36 |
|
| 37 |
-
# Actual model performance metrics
|
| 38 |
-
MODEL_METRICS = {
|
| 39 |
-
'segmentation': {
|
| 40 |
-
'dice': 0.6212,
|
| 41 |
-
'iou': 0.4506,
|
| 42 |
-
'precision': 0.7077,
|
| 43 |
-
'recall': 0.5536
|
| 44 |
-
},
|
| 45 |
-
'classification': {
|
| 46 |
-
'overall_accuracy': 0.8897,
|
| 47 |
-
'per_class': {
|
| 48 |
-
'copy_move': 0.92,
|
| 49 |
-
'splicing': 0.85,
|
| 50 |
-
'generation': 0.90
|
| 51 |
-
}
|
| 52 |
-
}
|
| 53 |
-
}
|
| 54 |
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
value=value * 100,
|
| 61 |
-
domain={'x': [0, 1], 'y': [0, 1]},
|
| 62 |
-
title={'text': title, 'font': {'size': 14}},
|
| 63 |
-
number={'suffix': '%', 'font': {'size': 24}},
|
| 64 |
-
gauge={
|
| 65 |
-
'axis': {'range': [0, 100], 'tickwidth': 1},
|
| 66 |
-
'bar': {'color': '#4169E1', 'thickness': 0.7},
|
| 67 |
-
'bgcolor': 'rgba(0,0,0,0)',
|
| 68 |
-
'borderwidth': 0,
|
| 69 |
-
'steps': [
|
| 70 |
-
{'range': [0, 50], 'color': 'rgba(217, 83, 79, 0.1)'},
|
| 71 |
-
{'range': [50, 75], 'color': 'rgba(240, 173, 78, 0.1)'},
|
| 72 |
-
{'range': [75, 100], 'color': 'rgba(92, 184, 92, 0.1)'}
|
| 73 |
-
]
|
| 74 |
-
}
|
| 75 |
-
))
|
| 76 |
-
|
| 77 |
-
fig.update_layout(
|
| 78 |
-
paper_bgcolor='rgba(0,0,0,0)',
|
| 79 |
-
plot_bgcolor='rgba(0,0,0,0)',
|
| 80 |
-
height=200,
|
| 81 |
-
margin=dict(l=20, r=20, t=40, b=20)
|
| 82 |
-
)
|
| 83 |
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
def create_detection_metrics_gauge(avg_confidence: float, iou: float, precision: float, recall: float, num_detections: int) -> go.Figure:
|
| 88 |
-
"""Create a high-fidelity radial bar chart (concentric rings)"""
|
| 89 |
-
|
| 90 |
-
# Calculate percentages (0-100)
|
| 91 |
-
metrics = [
|
| 92 |
-
{'name': 'Confidence', 'val': avg_confidence * 100 if num_detections > 0 else 0, 'color': '#4169E1', 'base': 80},
|
| 93 |
-
{'name': 'Precision', 'val': precision * 100, 'color': '#5cb85c', 'base': 60},
|
| 94 |
-
{'name': 'Recall', 'val': recall * 100, 'color': '#f0ad4e', 'base': 40},
|
| 95 |
-
{'name': 'IoU', 'val': iou * 100, 'color': '#d9534f', 'base': 20}
|
| 96 |
-
]
|
| 97 |
-
|
| 98 |
-
fig = go.Figure()
|
| 99 |
-
|
| 100 |
-
for m in metrics:
|
| 101 |
-
# 1. Add background track (faint gray ring)
|
| 102 |
-
fig.add_trace(go.Barpolar(
|
| 103 |
-
r=[15],
|
| 104 |
-
theta=[180],
|
| 105 |
-
width=[360],
|
| 106 |
-
base=m['base'],
|
| 107 |
-
marker_color='rgba(128,128,128,0.1)',
|
| 108 |
-
hoverinfo='none',
|
| 109 |
-
showlegend=False
|
| 110 |
-
))
|
| 111 |
-
|
| 112 |
-
# 2. Add the actual metric bar (the colored arc)
|
| 113 |
-
# 100% = 360 degrees
|
| 114 |
-
angle_width = m['val'] * 3.6
|
| 115 |
-
fig.add_trace(go.Barpolar(
|
| 116 |
-
r=[15],
|
| 117 |
-
theta=[angle_width / 2],
|
| 118 |
-
width=[angle_width],
|
| 119 |
-
base=m['base'],
|
| 120 |
-
name=f"{m['name']}: {m['val']:.1f}%",
|
| 121 |
-
marker_color=m['color'],
|
| 122 |
-
marker_line_width=0,
|
| 123 |
-
hoverinfo='name'
|
| 124 |
-
))
|
| 125 |
-
|
| 126 |
-
fig.update_layout(
|
| 127 |
-
polar=dict(
|
| 128 |
-
hole=0.1,
|
| 129 |
-
radialaxis=dict(visible=False, range=[0, 100]),
|
| 130 |
-
angularaxis=dict(
|
| 131 |
-
rotation=90, # Start at 12 o'clock
|
| 132 |
-
direction='clockwise', # Go clockwise
|
| 133 |
-
gridcolor='rgba(128,128,128,0.2)',
|
| 134 |
-
tickmode='array',
|
| 135 |
-
tickvals=[0, 90, 180, 270],
|
| 136 |
-
ticktext=['0%', '25%', '50%', '75%'],
|
| 137 |
-
showticklabels=True,
|
| 138 |
-
tickfont=dict(size=12, color='#888')
|
| 139 |
-
),
|
| 140 |
-
bgcolor='rgba(0,0,0,0)'
|
| 141 |
-
),
|
| 142 |
-
showlegend=True,
|
| 143 |
-
legend=dict(
|
| 144 |
-
orientation="v",
|
| 145 |
-
yanchor="middle",
|
| 146 |
-
y=0.5,
|
| 147 |
-
xanchor="left",
|
| 148 |
-
x=1.1,
|
| 149 |
-
font=dict(size=14, color='white'),
|
| 150 |
-
itemwidth=30
|
| 151 |
-
),
|
| 152 |
-
paper_bgcolor='rgba(0,0,0,0)',
|
| 153 |
-
plot_bgcolor='rgba(0,0,0,0)',
|
| 154 |
-
height=450,
|
| 155 |
-
margin=dict(l=60, r=180, t=40, b=40)
|
| 156 |
-
)
|
| 157 |
-
|
| 158 |
-
return fig
|
| 159 |
-
|
| 160 |
-
|
| 161 |
-
class ForgeryDetector:
|
| 162 |
-
"""Main forgery detection pipeline"""
|
| 163 |
-
|
| 164 |
-
def __init__(self):
|
| 165 |
-
print("Loading models...")
|
| 166 |
-
|
| 167 |
-
# Load config
|
| 168 |
-
self.config = get_config('config.yaml')
|
| 169 |
-
self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
| 170 |
|
| 171 |
-
|
| 172 |
-
|
| 173 |
-
|
| 174 |
-
|
| 175 |
-
self.
|
|
|
|
| 176 |
|
| 177 |
-
#
|
| 178 |
-
self.
|
| 179 |
-
self.
|
|
|
|
| 180 |
|
| 181 |
-
#
|
| 182 |
-
self.
|
| 183 |
-
self.augmentation = DatasetAwareAugmentation(self.config, 'doctamper', is_training=False)
|
| 184 |
-
self.mask_refiner = get_mask_refiner(self.config)
|
| 185 |
-
self.region_extractor = get_region_extractor(self.config)
|
| 186 |
-
self.feature_extractor = get_feature_extractor(self.config, is_text_document=True)
|
| 187 |
|
| 188 |
-
print("
|
|
|
|
| 189 |
|
| 190 |
-
def
|
|
|
|
|
|
|
| 191 |
"""
|
| 192 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 193 |
|
| 194 |
Returns:
|
| 195 |
-
|
| 196 |
-
overlay_image: Image with detection overlay
|
| 197 |
-
gauge_dice: Dice score gauge
|
| 198 |
-
gauge_accuracy: Accuracy gauge
|
| 199 |
-
results_html: Detection results as HTML
|
| 200 |
"""
|
| 201 |
-
#
|
| 202 |
-
|
| 203 |
-
if image.lower().endswith(('.doc', '.docx')):
|
| 204 |
-
# Handle Word documents - multiple fallback strategies
|
| 205 |
-
import tempfile
|
| 206 |
-
import os
|
| 207 |
-
import subprocess
|
| 208 |
-
|
| 209 |
-
temp_pdf = None
|
| 210 |
-
try:
|
| 211 |
-
# Strategy 1: Try docx2pdf (Windows with MS Word)
|
| 212 |
-
try:
|
| 213 |
-
from docx2pdf import convert
|
| 214 |
-
temp_pdf = tempfile.NamedTemporaryFile(delete=False, suffix='.pdf')
|
| 215 |
-
temp_pdf.close()
|
| 216 |
-
convert(image, temp_pdf.name)
|
| 217 |
-
pdf_path = temp_pdf.name
|
| 218 |
-
except Exception as e1:
|
| 219 |
-
# Strategy 2: Try LibreOffice (Linux/Mac)
|
| 220 |
-
try:
|
| 221 |
-
temp_pdf = tempfile.NamedTemporaryFile(delete=False, suffix='.pdf')
|
| 222 |
-
temp_pdf.close()
|
| 223 |
-
subprocess.run([
|
| 224 |
-
'libreoffice', '--headless', '--convert-to', 'pdf',
|
| 225 |
-
'--outdir', os.path.dirname(temp_pdf.name),
|
| 226 |
-
image
|
| 227 |
-
], check=True, capture_output=True)
|
| 228 |
-
|
| 229 |
-
# LibreOffice creates file with original name + .pdf
|
| 230 |
-
base_name = os.path.splitext(os.path.basename(image))[0]
|
| 231 |
-
generated_pdf = os.path.join(os.path.dirname(temp_pdf.name), f"{base_name}.pdf")
|
| 232 |
-
|
| 233 |
-
if os.path.exists(generated_pdf):
|
| 234 |
-
os.rename(generated_pdf, temp_pdf.name)
|
| 235 |
-
pdf_path = temp_pdf.name
|
| 236 |
-
else:
|
| 237 |
-
raise Exception("LibreOffice conversion failed")
|
| 238 |
-
except Exception as e2:
|
| 239 |
-
# Strategy 3: Extract text and create simple image
|
| 240 |
-
from docx import Document
|
| 241 |
-
doc = Document(image)
|
| 242 |
-
|
| 243 |
-
# Extract text
|
| 244 |
-
text_lines = []
|
| 245 |
-
for para in doc.paragraphs[:40]: # First 40 paragraphs
|
| 246 |
-
if para.text.strip():
|
| 247 |
-
text_lines.append(para.text[:100]) # Max 100 chars per line
|
| 248 |
-
|
| 249 |
-
# Create image with text
|
| 250 |
-
img_height = 1400
|
| 251 |
-
img_width = 1000
|
| 252 |
-
image = np.ones((img_height, img_width, 3), dtype=np.uint8) * 255
|
| 253 |
-
|
| 254 |
-
y_offset = 60
|
| 255 |
-
for line in text_lines[:35]:
|
| 256 |
-
cv2.putText(image, line, (40, y_offset),
|
| 257 |
-
cv2.FONT_HERSHEY_SIMPLEX, 0.45, (0, 0, 0), 1, cv2.LINE_AA)
|
| 258 |
-
y_offset += 35
|
| 259 |
-
|
| 260 |
-
# Skip to end - image is ready
|
| 261 |
-
pdf_path = None
|
| 262 |
-
|
| 263 |
-
# If we got a PDF, convert it to image
|
| 264 |
-
if pdf_path and os.path.exists(pdf_path):
|
| 265 |
-
import fitz
|
| 266 |
-
pdf_document = fitz.open(pdf_path)
|
| 267 |
-
page = pdf_document[0]
|
| 268 |
-
pix = page.get_pixmap(matrix=fitz.Matrix(2, 2))
|
| 269 |
-
image = np.frombuffer(pix.samples, dtype=np.uint8).reshape(pix.height, pix.width, pix.n)
|
| 270 |
-
if pix.n == 4:
|
| 271 |
-
image = cv2.cvtColor(image, cv2.COLOR_RGBA2RGB)
|
| 272 |
-
pdf_document.close()
|
| 273 |
-
os.unlink(pdf_path)
|
| 274 |
-
|
| 275 |
-
except Exception as e:
|
| 276 |
-
raise ValueError(f"Could not process Word document. Please convert to PDF or image first. Error: {str(e)}")
|
| 277 |
-
finally:
|
| 278 |
-
# Clean up temp file if it exists
|
| 279 |
-
if temp_pdf and os.path.exists(temp_pdf.name):
|
| 280 |
-
try:
|
| 281 |
-
os.unlink(temp_pdf.name)
|
| 282 |
-
except:
|
| 283 |
-
pass
|
| 284 |
-
|
| 285 |
-
elif image.lower().endswith('.pdf'):
|
| 286 |
-
# Handle PDF files
|
| 287 |
-
import fitz # PyMuPDF
|
| 288 |
-
pdf_document = fitz.open(image)
|
| 289 |
-
page = pdf_document[0]
|
| 290 |
-
pix = page.get_pixmap(matrix=fitz.Matrix(2, 2))
|
| 291 |
-
image = np.frombuffer(pix.samples, dtype=np.uint8).reshape(pix.height, pix.width, pix.n)
|
| 292 |
-
if pix.n == 4:
|
| 293 |
-
image = cv2.cvtColor(image, cv2.COLOR_RGBA2RGB)
|
| 294 |
-
pdf_document.close()
|
| 295 |
-
else:
|
| 296 |
-
# Load image file
|
| 297 |
-
image = Image.open(image)
|
| 298 |
-
image = np.array(image)
|
| 299 |
-
|
| 300 |
-
# Convert PIL to numpy
|
| 301 |
-
if isinstance(image, Image.Image):
|
| 302 |
-
image = np.array(image)
|
| 303 |
-
|
| 304 |
-
# Convert to RGB
|
| 305 |
-
if len(image.shape) == 2:
|
| 306 |
-
image = cv2.cvtColor(image, cv2.COLOR_GRAY2RGB)
|
| 307 |
-
elif image.shape[2] == 4:
|
| 308 |
-
image = cv2.cvtColor(image, cv2.COLOR_RGBA2RGB)
|
| 309 |
-
|
| 310 |
-
original_image = image.copy()
|
| 311 |
-
|
| 312 |
-
# Preprocess
|
| 313 |
-
preprocessed, _ = self.preprocessor(image, None)
|
| 314 |
|
| 315 |
-
#
|
| 316 |
-
|
| 317 |
-
|
| 318 |
-
|
| 319 |
-
|
| 320 |
-
|
| 321 |
-
logits, decoder_features = self.model(image_tensor)
|
| 322 |
-
prob_map = torch.sigmoid(logits).cpu().numpy()[0, 0]
|
| 323 |
|
| 324 |
-
#
|
| 325 |
-
|
| 326 |
-
|
| 327 |
-
(
|
| 328 |
-
interpolation=cv2.INTER_LINEAR
|
| 329 |
)
|
|
|
|
| 330 |
|
| 331 |
-
#
|
| 332 |
-
binary_mask =
|
| 333 |
-
refined_mask = self.mask_refiner.refine(prob_map_resized, original_size=original_image.shape[:2])
|
| 334 |
|
| 335 |
-
#
|
| 336 |
-
if
|
| 337 |
-
|
| 338 |
-
|
| 339 |
-
(
|
| 340 |
interpolation=cv2.INTER_NEAREST
|
| 341 |
)
|
| 342 |
|
| 343 |
-
|
| 344 |
-
if prob_map_resized.shape != refined_mask.shape:
|
| 345 |
-
prob_map_resized = cv2.resize(
|
| 346 |
-
prob_map_resized,
|
| 347 |
-
(refined_mask.shape[1], refined_mask.shape[0]),
|
| 348 |
-
interpolation=cv2.INTER_LINEAR
|
| 349 |
-
)
|
| 350 |
-
|
| 351 |
-
# Extract regions
|
| 352 |
-
regions = self.region_extractor.extract(refined_mask, prob_map_resized, original_image)
|
| 353 |
-
|
| 354 |
-
# Classify regions
|
| 355 |
-
results = []
|
| 356 |
-
for region in regions:
|
| 357 |
-
# Get decoder features and handle shape
|
| 358 |
-
df = decoder_features[0].cpu() # Get first decoder feature
|
| 359 |
-
|
| 360 |
-
# Remove batch dimension if present: [1, C, H, W] -> [C, H, W]
|
| 361 |
-
if df.ndim == 4:
|
| 362 |
-
df = df.squeeze(0)
|
| 363 |
-
|
| 364 |
-
# Now df should be [C, H, W]
|
| 365 |
-
_, fh, fw = df.shape
|
| 366 |
-
|
| 367 |
-
region_mask = region['region_mask']
|
| 368 |
-
if region_mask.shape != (fh, fw):
|
| 369 |
-
region_mask = cv2.resize(
|
| 370 |
-
region_mask.astype(np.uint8),
|
| 371 |
-
(fw, fh),
|
| 372 |
-
interpolation=cv2.INTER_NEAREST
|
| 373 |
-
)
|
| 374 |
-
|
| 375 |
-
region_mask = region_mask.astype(bool)
|
| 376 |
-
|
| 377 |
-
# Extract features
|
| 378 |
-
features = self.feature_extractor.extract(
|
| 379 |
-
preprocessed,
|
| 380 |
-
region['region_mask'],
|
| 381 |
-
[f.cpu() for f in decoder_features]
|
| 382 |
-
)
|
| 383 |
-
|
| 384 |
-
# Reshape features to 2D array
|
| 385 |
-
if features.ndim == 1:
|
| 386 |
-
features = features.reshape(1, -1)
|
| 387 |
-
|
| 388 |
-
# Pad/truncate features to match classifier
|
| 389 |
-
expected_features = 526
|
| 390 |
-
current_features = features.shape[1]
|
| 391 |
-
if current_features < expected_features:
|
| 392 |
-
padding = np.zeros((features.shape[0], expected_features - current_features))
|
| 393 |
-
features = np.hstack([features, padding])
|
| 394 |
-
elif current_features > expected_features:
|
| 395 |
-
features = features[:, :expected_features]
|
| 396 |
-
|
| 397 |
-
# Classify
|
| 398 |
-
predictions, confidences = self.classifier.predict(features)
|
| 399 |
-
forgery_type = int(predictions[0])
|
| 400 |
-
confidence = float(confidences[0])
|
| 401 |
-
|
| 402 |
-
if confidence > 0.6:
|
| 403 |
-
results.append({
|
| 404 |
-
'region_id': region['region_id'],
|
| 405 |
-
'bounding_box': region['bounding_box'],
|
| 406 |
-
'forgery_type': CLASS_NAMES[forgery_type],
|
| 407 |
-
'confidence': confidence
|
| 408 |
-
})
|
| 409 |
-
|
| 410 |
-
# Create visualization
|
| 411 |
-
overlay = self._create_overlay(original_image, results)
|
| 412 |
-
|
| 413 |
-
# Calculate actual detection metrics from probability map and mask
|
| 414 |
-
num_detections = len(results)
|
| 415 |
-
avg_confidence = sum(r['confidence'] for r in results) / num_detections if num_detections > 0 else 0
|
| 416 |
-
|
| 417 |
-
# Calculate IoU, Precision, Recall from the refined mask and probability map
|
| 418 |
-
if num_detections > 0:
|
| 419 |
-
# Use resized prob_map to match refined_mask dimensions
|
| 420 |
-
high_conf_mask = (prob_map_resized > 0.7).astype(np.uint8)
|
| 421 |
-
predicted_positive = np.sum(refined_mask > 0)
|
| 422 |
-
high_conf_positive = np.sum(high_conf_mask > 0)
|
| 423 |
-
|
| 424 |
-
# Calculate intersection and union
|
| 425 |
-
intersection = np.sum((refined_mask > 0) & (high_conf_mask > 0))
|
| 426 |
-
union = np.sum((refined_mask > 0) | (high_conf_mask > 0))
|
| 427 |
-
|
| 428 |
-
# Calculate metrics
|
| 429 |
-
iou = intersection / union if union > 0 else 0
|
| 430 |
-
precision = intersection / predicted_positive if predicted_positive > 0 else 0
|
| 431 |
-
recall = intersection / high_conf_positive if high_conf_positive > 0 else 0
|
| 432 |
-
else:
|
| 433 |
-
# No detections - use zeros
|
| 434 |
-
iou = 0
|
| 435 |
-
precision = 0
|
| 436 |
-
recall = 0
|
| 437 |
-
|
| 438 |
-
# Create detection metrics gauge with actual values
|
| 439 |
-
metrics_gauge = create_detection_metrics_gauge(avg_confidence, iou, precision, recall, num_detections)
|
| 440 |
-
|
| 441 |
-
# Create HTML response
|
| 442 |
-
results_html = self._create_html_report(results)
|
| 443 |
-
|
| 444 |
-
return overlay, metrics_gauge, results_html
|
| 445 |
|
| 446 |
-
def
|
| 447 |
-
"""
|
| 448 |
-
|
| 449 |
|
| 450 |
-
|
| 451 |
-
|
| 452 |
-
x, y, w, h = bbox
|
| 453 |
-
|
| 454 |
-
forgery_type = result['forgery_type']
|
| 455 |
-
confidence = result['confidence']
|
| 456 |
-
|
| 457 |
-
# Get color
|
| 458 |
-
forgery_id = [k for k, v in CLASS_NAMES.items() if v == forgery_type][0]
|
| 459 |
-
color = CLASS_COLORS[forgery_id]
|
| 460 |
-
|
| 461 |
-
# Draw rectangle
|
| 462 |
-
cv2.rectangle(overlay, (x, y), (x+w, y+h), color, 2)
|
| 463 |
-
|
| 464 |
-
# Draw label
|
| 465 |
-
label = f"{forgery_type}: {confidence:.1%}"
|
| 466 |
-
font = cv2.FONT_HERSHEY_SIMPLEX
|
| 467 |
-
font_scale = 0.5
|
| 468 |
-
thickness = 1
|
| 469 |
-
(label_w, label_h), baseline = cv2.getTextSize(label, font, font_scale, thickness)
|
| 470 |
-
|
| 471 |
-
cv2.rectangle(overlay, (x, y-label_h-8), (x+label_w+4, y), color, -1)
|
| 472 |
-
cv2.putText(overlay, label, (x+2, y-4), font, font_scale, (255, 255, 255), thickness)
|
| 473 |
|
| 474 |
-
|
| 475 |
-
|
| 476 |
-
|
| 477 |
-
|
| 478 |
-
|
|
|
|
| 479 |
|
| 480 |
-
|
| 481 |
-
|
| 482 |
-
<div style='padding:12px; border:1px solid #5cb85c; border-radius:8px;'>
|
| 483 |
-
✓ <b>No forgery detected.</b><br>
|
| 484 |
-
The document appears to be authentic.
|
| 485 |
-
</div>
|
| 486 |
-
"""
|
| 487 |
|
| 488 |
-
#
|
| 489 |
-
|
| 490 |
-
type_counts = {}
|
| 491 |
-
for r in results:
|
| 492 |
-
ft = r['forgery_type']
|
| 493 |
-
type_counts[ft] = type_counts.get(ft, 0) + 1
|
| 494 |
|
| 495 |
-
|
| 496 |
-
|
| 497 |
-
|
| 498 |
-
|
| 499 |
-
<b>Summary:</b><br>
|
| 500 |
-
• Regions detected: {num_detections}<br>
|
| 501 |
-
• Average confidence: {avg_confidence*100:.1f}%<br><br>
|
| 502 |
|
| 503 |
-
|
| 504 |
-
|
| 505 |
|
| 506 |
-
|
| 507 |
-
forgery_type = result['forgery_type']
|
| 508 |
-
confidence = result['confidence']
|
| 509 |
-
bbox = result['bounding_box']
|
| 510 |
-
|
| 511 |
-
forgery_id = [k for k, v in CLASS_NAMES.items() if v == forgery_type][0]
|
| 512 |
-
color_rgb = CLASS_COLORS[forgery_id]
|
| 513 |
-
color_hex = f"#{color_rgb[0]:02x}{color_rgb[1]:02x}{color_rgb[2]:02x}"
|
| 514 |
-
|
| 515 |
-
html += f"""
|
| 516 |
-
<div style='margin:8px 0; padding:8px; border-left:3px solid {color_hex}; background:rgba(0,0,0,0.02);'>
|
| 517 |
-
<b>Region {i}:</b> {forgery_type} ({confidence*100:.1f}%)<br>
|
| 518 |
-
<small>Location: ({bbox[0]}, {bbox[1]}) | Size: {bbox[2]}×{bbox[3]}px</small>
|
| 519 |
-
</div>
|
| 520 |
-
"""
|
| 521 |
-
|
| 522 |
-
html += """
|
| 523 |
-
</div>
|
| 524 |
-
"""
|
| 525 |
-
|
| 526 |
-
return html
|
| 527 |
|
| 528 |
|
| 529 |
-
|
| 530 |
-
|
| 531 |
-
|
| 532 |
-
|
| 533 |
-
|
| 534 |
-
|
| 535 |
-
|
| 536 |
-
|
| 537 |
-
|
| 538 |
|
| 539 |
-
|
| 540 |
-
|
| 541 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 542 |
|
| 543 |
-
|
| 544 |
-
|
|
|
|
|
|
|
| 545 |
|
| 546 |
-
|
| 547 |
-
|
| 548 |
-
except Exception as e:
|
| 549 |
-
import traceback
|
| 550 |
-
error_details = traceback.format_exc()
|
| 551 |
-
print(f"Error: {error_details}")
|
| 552 |
-
error_html = f"""
|
| 553 |
-
<div style='padding:12px; border:1px solid #d9534f; border-radius:8px;'>
|
| 554 |
-
❌ <b>Error:</b> {str(e)}
|
| 555 |
-
</div>
|
| 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 |
-
gr.Markdown("---")
|
| 582 |
-
|
| 583 |
-
with gr.Row():
|
| 584 |
-
with gr.Column(scale=1):
|
| 585 |
-
gr.Markdown("### Upload Document")
|
| 586 |
-
|
| 587 |
-
with gr.Tabs():
|
| 588 |
-
with gr.Tab("📤 Upload File"):
|
| 589 |
-
input_file = gr.File(
|
| 590 |
-
label="Upload Image, PDF, or Document",
|
| 591 |
-
file_types=["image", ".pdf", ".doc", ".docx"],
|
| 592 |
-
type="filepath"
|
| 593 |
-
)
|
| 594 |
-
|
| 595 |
-
with gr.Tab("📷 Webcam"):
|
| 596 |
-
input_webcam = gr.Image(
|
| 597 |
-
label="Capture from Webcam",
|
| 598 |
-
type="filepath",
|
| 599 |
-
sources=["webcam"]
|
| 600 |
-
)
|
| 601 |
-
|
| 602 |
-
with gr.Row():
|
| 603 |
-
clear_btn = gr.Button("🧹 Clear", elem_classes="clear-btn")
|
| 604 |
-
analyze_btn = gr.Button("🔍 Analyze", elem_classes="predict-btn")
|
| 605 |
|
| 606 |
-
|
| 607 |
-
|
| 608 |
-
gr.HTML(
|
| 609 |
-
"""
|
| 610 |
-
<div style='padding:16px; border:1px solid #ccc; border-radius:8px; background:var(--background-fill-primary);'>
|
| 611 |
-
<p style='margin-top:0;'><b>Supported formats:</b></p>
|
| 612 |
-
<ul style='margin:8px 0; padding-left:20px;'>
|
| 613 |
-
<li>Images: JPG, PNG, BMP, TIFF, WebP</li>
|
| 614 |
-
<li>PDF: First page analyzed</li>
|
| 615 |
-
</ul>
|
| 616 |
-
|
| 617 |
-
<p style='margin-bottom:4px;'><b>Forgery types:</b></p>
|
| 618 |
-
<ul style='margin:8px 0; padding-left:20px;'>
|
| 619 |
-
<li style='color:#d9534f;'><b>Copy-Move:</b> <span style='color:inherit;'>Duplicated regions</span></li>
|
| 620 |
-
<li style='color:#4169E1;'><b>Splicing:</b> <span style='color:inherit;'>Mixed sources</span></li>
|
| 621 |
-
<li style='color:#5cb85c;'><b>Text Substitution:</b> <span style='color:inherit;'>Modified text</span></li>
|
| 622 |
-
</ul>
|
| 623 |
-
</div>
|
| 624 |
-
"""
|
| 625 |
-
)
|
| 626 |
|
| 627 |
-
|
| 628 |
-
|
| 629 |
-
output_image = gr.Image(label="Detected Forgeries", type="numpy")
|
| 630 |
-
|
| 631 |
-
gr.Markdown("---")
|
| 632 |
-
|
| 633 |
-
with gr.Row():
|
| 634 |
-
with gr.Column(scale=1):
|
| 635 |
-
gr.Markdown("### Analysis Report")
|
| 636 |
-
output_html = gr.HTML(
|
| 637 |
-
value="<i>No analysis yet. Upload a document and click Analyze.</i>"
|
| 638 |
-
)
|
| 639 |
|
| 640 |
-
|
| 641 |
-
|
| 642 |
-
|
| 643 |
-
|
| 644 |
-
|
| 645 |
-
|
| 646 |
-
|
| 647 |
-
|
| 648 |
-
|
| 649 |
-
|
| 650 |
-
|
| 651 |
-
<div style='padding:12px; border:1px solid #444; border-radius:10px; background:var(--background-fill-primary);'>
|
| 652 |
-
<p style="margin:0 0 0px 0; font-size:1.05em;"><b>Localization:</b> MobileNetV3-Small + UNet</p>
|
| 653 |
-
<p style='margin:0 20px 5px 0; margin-left:0.5cm; font-size:0.9em; opacity:0.85;'>Dice: 62.12% | IoU: 45.06% | Precision: 70.77% | Recall: 55.36%</p>
|
| 654 |
-
|
| 655 |
-
<p style="margin:0 0 0 0; font-size:1.05em;"><b>Classification:</b> LightGBM with 526 features</p>
|
| 656 |
-
<p style="margin:0 20px 0 0; margin-left:0.5cm; font-size:0.9em; opacity:0.85;">Train Accuracy: 90.53% | Val Accuracy: 88.97%</p>
|
| 657 |
|
| 658 |
-
|
| 659 |
-
|
| 660 |
-
|
| 661 |
-
|
| 662 |
-
|
| 663 |
-
with gr.Column(scale=1):
|
| 664 |
-
gr.Markdown("### Model Performance")
|
| 665 |
-
gr.HTML(
|
| 666 |
-
f"""
|
| 667 |
-
<div style='padding:12px; border:1px solid #444; border-radius:10px; background:var(--background-fill-primary);'>
|
| 668 |
-
<p style='margin-top:0; margin-bottom:12px;'><b>Trained Model Performance:</b></p>
|
| 669 |
-
|
| 670 |
-
<b>Segmentation Dice: {MODEL_METRICS['segmentation']['dice']*100:.2f}%</b>
|
| 671 |
-
<div style='width:100%; background:#333; height:12px; border-radius:6px; margin-bottom:12px;'>
|
| 672 |
-
<div style='width:{MODEL_METRICS['segmentation']['dice']*100:.1f}%; background:#4169E1; height:12px; border-radius:6px;'></div>
|
| 673 |
-
</div>
|
| 674 |
-
|
| 675 |
-
<b>Classification Accuracy: {MODEL_METRICS['classification']['overall_accuracy']*100:.2f}%</b>
|
| 676 |
-
<div style='width:100%; background:#333; height:12px; border-radius:6px;'>
|
| 677 |
-
<div style='width:{MODEL_METRICS['classification']['overall_accuracy']*100:.1f}%; background:#5cb85c; height:12px; border-radius:6px;'></div>
|
| 678 |
-
</div>
|
| 679 |
-
</div>
|
| 680 |
-
"""
|
| 681 |
-
)
|
| 682 |
-
|
| 683 |
-
# Event handlers
|
| 684 |
-
analyze_btn.click(
|
| 685 |
-
fn=detect_forgery,
|
| 686 |
-
inputs=[input_file, input_webcam],
|
| 687 |
-
outputs=[output_image, metrics_gauge, output_html]
|
| 688 |
-
)
|
| 689 |
-
|
| 690 |
-
clear_btn.click(
|
| 691 |
-
fn=lambda: (None, None, None, None, "<i>No analysis yet. Upload a document and click Analyze.</i>"),
|
| 692 |
-
inputs=None,
|
| 693 |
-
outputs=[input_file, input_webcam, output_image, metrics_gauge, output_html]
|
| 694 |
-
)
|
| 695 |
|
| 696 |
|
| 697 |
-
|
| 698 |
-
|
|
|
|
|
|
| 1 |
"""
|
| 2 |
+
Mask refinement and region extraction
|
| 3 |
+
Implements Critical Fix #3: Adaptive Mask Refinement Thresholds
|
|
|
|
| 4 |
"""
|
| 5 |
|
|
|
|
|
|
|
| 6 |
import cv2
|
| 7 |
import numpy as np
|
| 8 |
+
from typing import List, Tuple, Dict, Optional
|
| 9 |
+
from scipy import ndimage
|
| 10 |
+
from skimage.measure import label, regionprops
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 11 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 12 |
|
| 13 |
+
class MaskRefiner:
|
| 14 |
+
"""
|
| 15 |
+
Mask refinement with adaptive thresholds
|
| 16 |
+
Implements Critical Fix #3: Dataset-specific minimum region areas
|
| 17 |
+
"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 18 |
|
| 19 |
+
def __init__(self, config, dataset_name: str = 'default'):
|
| 20 |
+
"""
|
| 21 |
+
Initialize mask refiner
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 22 |
|
| 23 |
+
Args:
|
| 24 |
+
config: Configuration object
|
| 25 |
+
dataset_name: Dataset name for adaptive thresholds
|
| 26 |
+
"""
|
| 27 |
+
self.config = config
|
| 28 |
+
self.dataset_name = dataset_name
|
| 29 |
|
| 30 |
+
# Get mask refinement parameters
|
| 31 |
+
self.threshold = config.get('mask_refinement.threshold', 0.5)
|
| 32 |
+
self.closing_kernel = config.get('mask_refinement.morphology.closing_kernel', 5)
|
| 33 |
+
self.opening_kernel = config.get('mask_refinement.morphology.opening_kernel', 3)
|
| 34 |
|
| 35 |
+
# Critical Fix #3: Adaptive thresholds per dataset
|
| 36 |
+
self.min_region_area = config.get_min_region_area(dataset_name)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 37 |
|
| 38 |
+
print(f"MaskRefiner initialized for {dataset_name}")
|
| 39 |
+
print(f"Min region area: {self.min_region_area * 100:.2f}%")
|
| 40 |
|
| 41 |
+
def refine(self,
|
| 42 |
+
probability_map: np.ndarray,
|
| 43 |
+
original_size: Tuple[int, int] = None) -> np.ndarray:
|
| 44 |
"""
|
| 45 |
+
Refine probability map to binary mask
|
| 46 |
+
|
| 47 |
+
Args:
|
| 48 |
+
probability_map: Forgery probability map (H, W), values [0, 1]
|
| 49 |
+
original_size: Optional (H, W) to resize mask back to original
|
| 50 |
|
| 51 |
Returns:
|
| 52 |
+
Refined binary mask (H, W)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 53 |
"""
|
| 54 |
+
# Threshold to binary
|
| 55 |
+
binary_mask = (probability_map > self.threshold).astype(np.uint8)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 56 |
|
| 57 |
+
# Morphological closing (fill broken strokes)
|
| 58 |
+
closing_kernel = cv2.getStructuringElement(
|
| 59 |
+
cv2.MORPH_RECT,
|
| 60 |
+
(self.closing_kernel, self.closing_kernel)
|
| 61 |
+
)
|
| 62 |
+
binary_mask = cv2.morphologyEx(binary_mask, cv2.MORPH_CLOSE, closing_kernel)
|
|
|
|
|
|
|
| 63 |
|
| 64 |
+
# Morphological opening (remove isolated noise)
|
| 65 |
+
opening_kernel = cv2.getStructuringElement(
|
| 66 |
+
cv2.MORPH_RECT,
|
| 67 |
+
(self.opening_kernel, self.opening_kernel)
|
|
|
|
| 68 |
)
|
| 69 |
+
binary_mask = cv2.morphologyEx(binary_mask, cv2.MORPH_OPEN, opening_kernel)
|
| 70 |
|
| 71 |
+
# Critical Fix #3: Remove small regions with adaptive threshold
|
| 72 |
+
binary_mask = self._remove_small_regions(binary_mask)
|
|
|
|
| 73 |
|
| 74 |
+
# Resize to original size if provided
|
| 75 |
+
if original_size is not None:
|
| 76 |
+
binary_mask = cv2.resize(
|
| 77 |
+
binary_mask,
|
| 78 |
+
(original_size[1], original_size[0]), # cv2 uses (W, H)
|
| 79 |
interpolation=cv2.INTER_NEAREST
|
| 80 |
)
|
| 81 |
|
| 82 |
+
return binary_mask
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 83 |
|
| 84 |
+
def _remove_small_regions(self, mask: np.ndarray) -> np.ndarray:
|
| 85 |
+
"""
|
| 86 |
+
Remove regions smaller than minimum area threshold
|
| 87 |
|
| 88 |
+
Args:
|
| 89 |
+
mask: Binary mask (H, W)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 90 |
|
| 91 |
+
Returns:
|
| 92 |
+
Filtered mask
|
| 93 |
+
"""
|
| 94 |
+
# Calculate minimum pixel count
|
| 95 |
+
image_area = mask.shape[0] * mask.shape[1]
|
| 96 |
+
min_pixels = int(image_area * self.min_region_area)
|
| 97 |
|
| 98 |
+
# Label connected components
|
| 99 |
+
labeled_mask, num_features = ndimage.label(mask)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 100 |
|
| 101 |
+
# Keep only large enough regions
|
| 102 |
+
filtered_mask = np.zeros_like(mask)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 103 |
|
| 104 |
+
for region_id in range(1, num_features + 1):
|
| 105 |
+
region_mask = (labeled_mask == region_id)
|
| 106 |
+
region_area = region_mask.sum()
|
|
|
|
|
|
|
|
|
|
|
|
|
| 107 |
|
| 108 |
+
if region_area >= min_pixels:
|
| 109 |
+
filtered_mask[region_mask] = 1
|
| 110 |
|
| 111 |
+
return filtered_mask
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 112 |
|
| 113 |
|
| 114 |
+
class RegionExtractor:
|
| 115 |
+
"""
|
| 116 |
+
Extract individual regions from binary mask
|
| 117 |
+
Implements Critical Fix #4: Region Confidence Aggregation
|
| 118 |
+
"""
|
| 119 |
+
|
| 120 |
+
def __init__(self, config, dataset_name: str = 'default'):
|
| 121 |
+
"""
|
| 122 |
+
Initialize region extractor
|
| 123 |
|
| 124 |
+
Args:
|
| 125 |
+
config: Configuration object
|
| 126 |
+
dataset_name: Dataset name
|
| 127 |
+
"""
|
| 128 |
+
self.config = config
|
| 129 |
+
self.dataset_name = dataset_name
|
| 130 |
+
self.min_region_area = config.get_min_region_area(dataset_name)
|
| 131 |
+
|
| 132 |
+
def extract(self,
|
| 133 |
+
binary_mask: np.ndarray,
|
| 134 |
+
probability_map: np.ndarray,
|
| 135 |
+
original_image: np.ndarray) -> List[Dict]:
|
| 136 |
+
"""
|
| 137 |
+
Extract regions from binary mask
|
| 138 |
|
| 139 |
+
Args:
|
| 140 |
+
binary_mask: Refined binary mask (H, W)
|
| 141 |
+
probability_map: Original probability map (H, W)
|
| 142 |
+
original_image: Original image (H, W, 3)
|
| 143 |
|
| 144 |
+
Returns:
|
| 145 |
+
List of region dictionaries with bounding box, mask, image, confidence
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 146 |
"""
|
| 147 |
+
regions = []
|
| 148 |
+
|
| 149 |
+
print(f"[REGION_EXTRACT] Input shapes:")
|
| 150 |
+
print(f" - binary_mask: {binary_mask.shape}")
|
| 151 |
+
print(f" - probability_map: {probability_map.shape}")
|
| 152 |
+
print(f" - original_image: {original_image.shape}")
|
| 153 |
+
|
| 154 |
+
# Safety check: Ensure probability_map and binary_mask have same dimensions
|
| 155 |
+
if probability_map.shape != binary_mask.shape:
|
| 156 |
+
print(f"[REGION_EXTRACT] WARNING: Shape mismatch! Resizing probability_map from {probability_map.shape} to {binary_mask.shape}")
|
| 157 |
+
import cv2
|
| 158 |
+
probability_map = cv2.resize(
|
| 159 |
+
probability_map,
|
| 160 |
+
(binary_mask.shape[1], binary_mask.shape[0]),
|
| 161 |
+
interpolation=cv2.INTER_LINEAR
|
| 162 |
+
)
|
| 163 |
+
print(f"[REGION_EXTRACT] After resize: probability_map shape = {probability_map.shape}")
|
| 164 |
+
|
| 165 |
+
# Connected component analysis (8-connectivity)
|
| 166 |
+
labeled_mask = label(binary_mask, connectivity=2)
|
| 167 |
+
props = regionprops(labeled_mask)
|
| 168 |
+
|
| 169 |
+
for region_id, prop in enumerate(props, start=1):
|
| 170 |
+
# Bounding box
|
| 171 |
+
y_min, x_min, y_max, x_max = prop.bbox
|
| 172 |
+
|
| 173 |
+
# Region mask
|
| 174 |
+
region_mask = (labeled_mask == region_id).astype(np.uint8)
|
| 175 |
+
|
| 176 |
+
# Cropped region image
|
| 177 |
+
region_image = original_image[y_min:y_max, x_min:x_max].copy()
|
| 178 |
+
region_mask_cropped = region_mask[y_min:y_max, x_min:x_max]
|
| 179 |
+
|
| 180 |
+
|
| 181 |
+
# Critical Fix #4: Region-level confidence aggregation
|
| 182 |
+
# Ensure region_mask and probability_map have same shape
|
| 183 |
+
if region_mask.shape != probability_map.shape:
|
| 184 |
+
import cv2
|
| 185 |
+
# Resize probability_map to match region_mask
|
| 186 |
+
probability_map = cv2.resize(
|
| 187 |
+
probability_map,
|
| 188 |
+
(region_mask.shape[1], region_mask.shape[0]),
|
| 189 |
+
interpolation=cv2.INTER_LINEAR
|
| 190 |
+
)
|
| 191 |
+
|
| 192 |
+
region_probs = probability_map[region_mask > 0]
|
| 193 |
+
region_confidence = float(np.mean(region_probs)) if len(region_probs) > 0 else 0.0
|
| 194 |
+
|
| 195 |
+
regions.append({
|
| 196 |
+
'region_id': region_id,
|
| 197 |
+
'bounding_box': [int(x_min), int(y_min),
|
| 198 |
+
int(x_max - x_min), int(y_max - y_min)],
|
| 199 |
+
'area': prop.area,
|
| 200 |
+
'centroid': (int(prop.centroid[1]), int(prop.centroid[0])),
|
| 201 |
+
'region_mask': region_mask,
|
| 202 |
+
'region_mask_cropped': region_mask_cropped,
|
| 203 |
+
'region_image': region_image,
|
| 204 |
+
'confidence': region_confidence,
|
| 205 |
+
'mask_probability_mean': region_confidence
|
| 206 |
+
})
|
| 207 |
+
|
| 208 |
+
return regions
|
| 209 |
+
|
| 210 |
+
def extract_for_casia(self,
|
| 211 |
+
binary_mask: np.ndarray,
|
| 212 |
+
probability_map: np.ndarray,
|
| 213 |
+
original_image: np.ndarray) -> List[Dict]:
|
| 214 |
"""
|
| 215 |
+
Critical Fix #6: CASIA handling - treat entire image as one region
|
| 216 |
+
|
| 217 |
+
Args:
|
| 218 |
+
binary_mask: Binary mask (may be empty for authentic images)
|
| 219 |
+
probability_map: Probability map
|
| 220 |
+
original_image: Original image
|
| 221 |
+
|
| 222 |
+
Returns:
|
| 223 |
+
Single region representing entire image
|
| 224 |
"""
|
| 225 |
+
h, w = original_image.shape[:2]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 226 |
|
| 227 |
+
# Create single region covering entire image
|
| 228 |
+
region_mask = np.ones((h, w), dtype=np.uint8)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 229 |
|
| 230 |
+
# Overall confidence from probability map
|
| 231 |
+
overall_confidence = float(np.mean(probability_map))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 232 |
|
| 233 |
+
return [{
|
| 234 |
+
'region_id': 1,
|
| 235 |
+
'bounding_box': [0, 0, w, h],
|
| 236 |
+
'area': h * w,
|
| 237 |
+
'centroid': (w // 2, h // 2),
|
| 238 |
+
'region_mask': region_mask,
|
| 239 |
+
'region_mask_cropped': region_mask,
|
| 240 |
+
'region_image': original_image,
|
| 241 |
+
'confidence': overall_confidence,
|
| 242 |
+
'mask_probability_mean': overall_confidence
|
| 243 |
+
}]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 244 |
|
| 245 |
+
|
| 246 |
+
def get_mask_refiner(config, dataset_name: str = 'default') -> MaskRefiner:
|
| 247 |
+
"""Factory function for mask refiner"""
|
| 248 |
+
return MaskRefiner(config, dataset_name)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 249 |
|
| 250 |
|
| 251 |
+
def get_region_extractor(config, dataset_name: str = 'default') -> RegionExtractor:
|
| 252 |
+
"""Factory function for region extractor"""
|
| 253 |
+
return RegionExtractor(config, dataset_name)
|