Update fen.py with HuggingFace Hub integration

fen.py

@@ -937,8 +937,8 @@ def run_onnx_yolo_detection_416(image, onnx_session):
             'black_bishop', 'black_rook', 'black_queen', 'black_king'
         ]
 
-        # Use lower confidence threshold for ONNX
-        conf_threshold = 0.1
+        # Use higher confidence threshold for better quality detections
+        conf_threshold = 0.4
 
         # Process each prediction
         for i, prediction in enumerate(predictions):
@@ -1010,9 +1010,59 @@ def run_onnx_yolo_detection_416(image, onnx_session):
                         'center': [center_x, center_y]
                     })
 
-                    print(f"  🔍 Detected {piece_name} at ({center_x:.1f}, {center_y:.1f}) conf: {final_conf:.3f}")
+    # Filter to keep only highest confidence detection for each board position
+    filtered_detections = filter_highest_confidence_detections(detections)
 
-    return detections
+    print(f"   📊 Filtered to {len(filtered_detections)} high-confidence detections")
+    for detection in filtered_detections:
+        piece_name = detection['piece']
+        center_x, center_y = detection['center']
+        conf = detection['confidence']
+        print(f"  🔍 Final {piece_name} at ({center_x:.1f}, {center_y:.1f}) conf: {conf:.3f}")
+
+    return filtered_detections
+
+def filter_highest_confidence_detections(detections, grid_size=64):
+    """Filter detections to keep only the highest confidence detection for each board position"""
+    print("🔍 Filtering detections to keep highest confidence per position...")
+
+    # Group detections by approximate board position
+    position_groups = {}
+
+    for detection in detections:
+        center_x, center_y = detection['center']
+
+        # Calculate approximate grid position (8x8 board)
+        grid_x = int(center_x // grid_size)
+        grid_y = int(center_y // grid_size)
+
+        # Clamp to valid board positions
+        grid_x = max(0, min(7, grid_x))
+        grid_y = max(0, min(7, grid_y))
+
+        position_key = f"{grid_x},{grid_y}"
+
+        if position_key not in position_groups:
+            position_groups[position_key] = []
+
+        position_groups[position_key].append(detection)
+
+    # Keep only the highest confidence detection for each position
+    filtered_detections = []
+
+    for position_key, group_detections in position_groups.items():
+        if len(group_detections) == 1:
+            # Only one detection at this position
+            filtered_detections.append(group_detections[0])
+        else:
+            # Multiple detections - keep the one with highest confidence
+            best_detection = max(group_detections, key=lambda d: d['confidence'])
+            filtered_detections.append(best_detection)
+
+            print(f"   📍 Position {position_key}: kept best of {len(group_detections)} detections (conf: {best_detection['confidence']:.3f})")
+
+    print(f"   ✅ Filtered from {len(detections)} to {len(filtered_detections)} detections")
+    return filtered_detections
 
 def map_pieces_to_board(detections, board_size=512):
     """Map pieces to 8x8 board with improved tolerance for pieces outside square boundaries"""