Enhance template matching in VideoEditor with multi-scale approach and adaptive thresholding

This commit introduces a multi-scale template matching technique to improve tracking accuracy by evaluating templates at various sizes. Additionally, it implements adaptive thresholding based on recent match confidences, allowing for more dynamic and responsive matching criteria. A new method for image preprocessing is also added to enhance contrast and reduce noise, further optimizing the template matching process.

croppa/main.py

@@ -880,7 +880,8 @@ class VideoEditor:
         # Template matching tracking
         self.template_matching_enabled = False
         self.tracking_template = None
-        self.template_region = None  # (x, y, w, h) in rotated frame coordinates
+        self.template_region = None
+        self.template_match_history = []  # Store recent match confidences for adaptive thresholding  # (x, y, w, h) in rotated frame coordinates
         self.template_selection_start = None
         self.template_selection_rect = None
 
@@ -2225,28 +2226,52 @@ class VideoEditor:
                 return None
         
         try:
-            # Convert to grayscale
-            if len(frame.shape) == 3:
-                gray_frame = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
-            else:
-                gray_frame = frame
-                
-            if len(self.tracking_template.shape) == 3:
-                gray_template = cv2.cvtColor(self.tracking_template, cv2.COLOR_BGR2GRAY)
-            else:
-                gray_template = self.tracking_template
+            # Apply image preprocessing for better template matching
+            gray_frame, gray_template = self._improve_template_matching(frame, self.tracking_template)
             
-            # Template matching
-            result = cv2.matchTemplate(gray_frame, gray_template, cv2.TM_CCOEFF_NORMED)
-            _, max_val, _, max_loc = cv2.minMaxLoc(result)
+            # Multi-scale template matching for better tracking
+            scales = [0.8, 0.9, 1.0, 1.1, 1.2]  # Different scales to try
+            best_match = None
+            best_confidence = 0.0
             
-            # Only accept matches above threshold
-            if max_val > 0.6:  # Adjust threshold as needed
-                # Get template center
+            for scale in scales:
+                # Resize template
                 template_h, template_w = gray_template.shape
-                center_x = max_loc[0] + template_w // 2
-                center_y = max_loc[1] + template_h // 2
-                return (center_x, center_y, max_val)
+                new_w = int(template_w * scale)
+                new_h = int(template_h * scale)
+                
+                if new_w <= 0 or new_h <= 0 or new_w > gray_frame.shape[1] or new_h > gray_frame.shape[0]:
+                    continue
+                    
+                scaled_template = cv2.resize(gray_template, (new_w, new_h))
+                
+                # Perform template matching
+                result = cv2.matchTemplate(gray_frame, scaled_template, cv2.TM_CCOEFF_NORMED)
+                _, max_val, _, max_loc = cv2.minMaxLoc(result)
+                
+                # Check if this is the best match so far
+                if max_val > best_confidence:
+                    best_confidence = max_val
+                    # Get center of template
+                    center_x = max_loc[0] + new_w // 2
+                    center_y = max_loc[1] + new_h // 2
+                    best_match = (center_x, center_y, max_val)
+            
+            # Adaptive thresholding based on recent match history
+            if len(self.template_match_history) > 0:
+                # Use average of recent matches as baseline
+                avg_confidence = sum(self.template_match_history[-10:]) / len(self.template_match_history[-10:])
+                threshold = max(0.3, avg_confidence * 0.8)  # 80% of recent average, minimum 0.3
+            else:
+                threshold = 0.5  # Default threshold
+            
+            # Only accept matches above adaptive threshold
+            if best_confidence > threshold:
+                # Store confidence for adaptive thresholding
+                self.template_match_history.append(best_confidence)
+                if len(self.template_match_history) > 20:  # Keep only last 20 matches
+                    self.template_match_history.pop(0)
+                return best_match
             else:
                 return None
                 
@@ -2284,6 +2309,33 @@ class VideoEditor:
         except Exception as e:
             print(f"Error recreating template: {e}")
             return False
+
+    def _improve_template_matching(self, frame, template):
+        """Apply image preprocessing to improve template matching"""
+        try:
+            # Convert to grayscale if needed
+            if len(frame.shape) == 3:
+                gray_frame = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
+            else:
+                gray_frame = frame
+                
+            if len(template.shape) == 3:
+                gray_template = cv2.cvtColor(template, cv2.COLOR_BGR2GRAY)
+            else:
+                gray_template = template
+            
+            # Apply histogram equalization for better contrast
+            gray_frame = cv2.equalizeHist(gray_frame)
+            gray_template = cv2.equalizeHist(gray_template)
+            
+            # Apply Gaussian blur to reduce noise
+            gray_frame = cv2.GaussianBlur(gray_frame, (3, 3), 0)
+            gray_template = cv2.GaussianBlur(gray_template, (3, 3), 0)
+            
+            return gray_frame, gray_template
+        except Exception as e:
+            print(f"Error improving template matching: {e}")
+            return frame, template
     
     def _set_template_from_region(self, screen_rect):
         """Set template from selected region"""