Update app.py

app.py

@@ -29,9 +29,9 @@ from src.training.classifier import ForgeryClassifier
 # Class names
 CLASS_NAMES = {0: 'Copy-Move', 1: 'Splicing', 2: 'Text Substitution'}
 CLASS_COLORS = {
-    0: (217, 83, 79),    # #d9534f - Muted red (Copy-Move)
-    1: (65, 105, 225),   # #4169E1 - Royal blue (Splicing)
-    2: (92, 184, 92)     # #5cb85c - Muted green (Text Substitution)
+    0: (217, 83, 79),    # #d9534f - Muted red
+    1: (92, 184, 92),    # #5cb85c - Muted green
+    2: (65, 105, 225)    # #4169E1 - Royal blue
 }
 
 # Actual model performance metrics
@@ -151,7 +151,7 @@ def create_detection_metrics_gauge(avg_confidence: float, iou: float, precision:
         ),
         paper_bgcolor='rgba(0,0,0,0)',
         plot_bgcolor='rgba(0,0,0,0)',
-        height=300,  # Reduced from 450
+        height=450,
         margin=dict(l=60, r=180, t=40, b=40)
     )
     
@@ -432,13 +432,10 @@ class ForgeryDetector:
 
             region_mask = region_mask.astype(bool)
             
-            # Extract features using tensor converted to numpy (matches training pipeline)
-            # Convert tensor back to numpy: (C, H, W) -> (H, W, C)
-            preprocessed_numpy = image_tensor[0].permute(1, 2, 0).cpu().numpy()
-            
-            # Pass region_mask directly - feature extractor handles resizing internally
+            # Extract features using ORIGINAL image (not preprocessed)
+            # This matches how the classifier was trained
             features = self.feature_extractor.extract(
-                preprocessed_numpy,
+                original_image / 255.0,  # Normalize to [0, 1]
                 region['region_mask'],
                 [f.cpu() for f in decoder_features]
             )