Update detect.py

detect.py

@@ -53,17 +53,17 @@ def _shape_probabilities_from_geometry(landmarks):
 
     targets = {
         "Round": {"lw": 1.05, "jaw": 0.88},
-        "Square": {"lw": 1.15, "jaw": 0.95},
-        "Oval": {"lw": 1.35, "jaw": 0.86},
-        "Oblong": {"lw": 1.60, "jaw": 0.86},
-        "Heart": {"lw": 1.35, "jaw": 0.75},
-        "Diamond": {"lw": 1.30, "jaw": 0.70},
+        "Square": {"lw": 1.12, "jaw": 0.95},
+        "Oval": {"lw": 1.25, "jaw": 0.86},
+        "Oblong": {"lw": 1.35, "jaw": 0.86},
+        "Heart": {"lw": 1.22, "jaw": 0.75},
+        "Diamond": {"lw": 1.20, "jaw": 0.70},
     }
 
     scores = {}
     for shape, target in targets.items():
-        lw_score = _gaussian_score(lw_ratio, target["lw"], 0.12)
-        jaw_score = _gaussian_score(jaw_ratio, target["jaw"], 0.08)
+        lw_score = _gaussian_score(lw_ratio, target["lw"], 0.1)
+        jaw_score = _gaussian_score(jaw_ratio, target["jaw"], 0.07)
         scores[shape] = lw_score * jaw_score
 
     total = sum(scores.values())
@@ -73,6 +73,20 @@ def _shape_probabilities_from_geometry(landmarks):
     probabilities = {k: round(v / total, 4) for k, v in scores.items()}
     return dict(sorted(probabilities.items(), key=lambda item: item[1], reverse=True))
 
+
+def _blend_probabilities(primary, secondary, alpha=0.55):
+    labels = set(primary.keys()) | set(secondary.keys())
+    blended = {}
+    for label in labels:
+        blended[label] = alpha * primary.get(label, 0) + (1 - alpha) * secondary.get(label, 0)
+    total = sum(blended.values())
+    if total <= 0:
+        return primary
+    blended = {k: round(v / total, 4) for k, v in blended.items()}
+    return dict(sorted(blended.items(), key=lambda item: item[1], reverse=True))
+
+
+
 def detect_face_shape(image_path):
     """
     Detects face shape using PIL and trained Mean Face model.
@@ -97,17 +111,24 @@ def detect_face_shape(image_path):
             img = img.convert("RGB")
             img = _crop_to_face(img)
 
+        geom_probs = None
         if isinstance(image_path, str):
             try:
                 landmarks = get_landmarks(image_path)
                 geom_probs = _shape_probabilities_from_geometry(landmarks)
-                if geom_probs:
-                    return geom_probs
             except Exception as e:
                 logger.info(f"Landmark-based detection failed, falling back: {e}")
 
         # Classify directly (classifier handles resizing/grayscale)
         shape_probs = classify_face_shape(img)
+
+        if geom_probs:
+            geom_values = list(geom_probs.values())
+            top = geom_values[0] if geom_values else 0
+            runner_up = geom_values[1] if len(geom_values) > 1 else 0
+            if top >= 0.45 and (top - runner_up) >= 0.08:
+                return geom_probs
+            return _blend_probabilities(geom_probs, shape_probs, alpha=0.55)
         
         # Get the top shape
         best_shape = list(shape_probs.keys())[0]