Fix Resize Error

main.py

@@ -68,26 +68,41 @@ def compute_semantic_similarity(img_a: Image.Image, img_b: Image.Image) -> float
     return round(score.item(), 4)
 
 def compute_structural_similarity(img_a: Image.Image, img_b: Image.Image):
+    # Convert to grayscale numpy arrays
     arr_a = np.array(img_a.convert('L'))
     arr_b = np.array(img_b.convert('L'))
     
-    edges_a = cv2.Canny(arr_a, 100, 200)
-    edges_b = cv2.Canny(arr_b, 100, 200)
-    
-    # 1. Create a kernel (a 5x5 pixel block)
-    kernel = np.ones((5,5), np.uint8)
+    # 1. Aggressive Gaussian Blur to eliminate stylistic texture
+    # A kernel size of (11, 11) is strong enough to blur out noise but keep main shapes
+    blur_a = cv2.GaussianBlur(arr_a, (11, 11), 0)
+    blur_b = cv2.GaussianBlur(arr_b, (11, 11), 0)
+    
+    # 2. Dynamic Auto-Canny Helper Function
+    def auto_canny(image, sigma=0.33):
+        v = np.median(image)
+        lower = int(max(0, (1.0 - sigma) * v))
+        upper = int(min(255, (1.0 + sigma) * v))
+        return cv2.Canny(image, lower, upper)
+        
+    # Extract structural edges using the blurred images and dynamic thresholds
+    edges_a = auto_canny(blur_a)
+    edges_b = auto_canny(blur_b)
     
-    # 2. Dilate (thicken) the edges
+    # 3. Dilate the edges to give them a "margin of error" for spatial overlap
+    kernel = np.ones((7,7), np.uint8) # Thicker kernel for better overlap
     edges_a_thick = cv2.dilate(edges_a, kernel, iterations=1)
     edges_b_thick = cv2.dilate(edges_b, kernel, iterations=1)
     
+    # Resize B to match A for matrix math
     edges_b_resized = cv2.resize(edges_b_thick, (edges_a_thick.shape[1], edges_a_thick.shape[0]))
     
-    # 3. Compute IoU on the thickened edges
+    # Calculate Intersection over Union (IoU)
     intersection = np.logical_and(edges_a_thick > 0, edges_b_resized > 0).sum()
     union = np.logical_or(edges_a_thick > 0, edges_b_resized > 0).sum()
     
     iou_score = intersection / union if union != 0 else 0.0
+    
+    # Return the clean (non-thickened) edges for a prettier UI visualization
     return round(iou_score, 4), Image.fromarray(edges_a), Image.fromarray(edges_b)
 
 def compute_patch_similarity(img_a: Image.Image, img_b: Image.Image):
@@ -95,8 +110,9 @@ def compute_patch_similarity(img_a: Image.Image, img_b: Image.Image):
     img_a_resized = img_a.resize(target_size)
     img_b_resized = img_b.resize(target_size)
     
-    inputs_a = dino_processor(images=img_a_resized, return_tensors="pt").to(device)
-    inputs_b = dino_processor(images=img_b_resized, return_tensors="pt").to(device)
+    # FIX: Add do_resize=False so it doesn't crush our 518x518 images back to 224x224
+    inputs_a = dino_processor(images=img_a_resized, return_tensors="pt", do_resize=False).to(device)
+    inputs_b = dino_processor(images=img_b_resized, return_tensors="pt", do_resize=False).to(device)
     
     with torch.no_grad():
         out_a = dino_model(**inputs_a)