Relaxed Image validation limits for mobile photos

main.py

@@ -219,9 +219,9 @@ async def predict(file: UploadFile = File(...)) -> dict:
     # ── 8. Validate confidence threshold ───────────────────────────────────
     top_confidence = predictions[0]["confidence"]
 
-    # Hard reject only truly uncertain predictions (lowered from 25 → 15 %
-    # because genuine skin photos of mild or early conditions can score low)
-    if top_confidence < 15.0:
+    # Hard reject only truly uncertain predictions (lowered from 25 -> 15 -> 5 %
+    # because genuine skin photos of mild or early conditions can score low, especially on smartphones)
+    if top_confidence < 5.0:
         raise HTTPException(
             status_code=422,
             detail={

model.py

@@ -288,7 +288,7 @@ def is_skin_image(image_array: np.ndarray) -> dict:
     skin_pixels = int(cv2.countNonZero(skin_mask))
     skin_pct    = (skin_pixels / total_pixels) * 100
 
-    MIN_SKIN_PCT = 15.0  # at least 15 % of the frame must be skin-coloured
+    MIN_SKIN_PCT = 3.0   # dramatically lowered to allow partial/small skin patches or phone crops
 
     if skin_pct < MIN_SKIN_PCT:
         return {
@@ -329,7 +329,7 @@ def is_skin_image(image_array: np.ndarray) -> dict:
 
     # The largest blob must cover ≥ 8 % of the total image
     largest_blob_pct = (max(large_blobs) / total_pixels) * 100
-    if largest_blob_pct < 8.0:
+    if largest_blob_pct < 3.0:
         return {
             "is_skin": False,
             "skin_percentage": round(skin_pct, 1),
@@ -348,8 +348,8 @@ def is_skin_image(image_array: np.ndarray) -> dict:
 
     if sat_skin.size > 0:
         sat_std = float(np.std(sat_skin))
-        # Paint / metal usually has sat_std < 18; real skin > 25
-        if sat_std < 15.0:
+        # Paint / metal usually has sat_std < 18; phone images can be heavily compressed and lose variance
+        if sat_std < 5.0:
             return {
                 "is_skin": False,
                 "skin_percentage": round(skin_pct, 1),
@@ -368,8 +368,8 @@ def is_skin_image(image_array: np.ndarray) -> dict:
 
     if skin_pixels > 0:
         edge_skin_ratio = edge_count / skin_pixels
-        # If edges outnumber skin pixels 1:1 this is almost certainly not skin
-        if edge_skin_ratio > 0.9:
+        # Relaxed edge density check to accommodate textured or noisy images
+        if edge_skin_ratio > 1.5:
             return {
                 "is_skin": False,
                 "skin_percentage": round(skin_pct, 1),
@@ -409,10 +409,9 @@ def check_image_quality(image_array: np.ndarray) -> dict:
     gray     = cv2.cvtColor(image_array, cv2.COLOR_RGB2GRAY)
     lap_var  = float(cv2.Laplacian(gray, cv2.CV_64F).var())
 
-    # Threshold calibrated so that genuine medical photos from modern phones
-    # (which often have high detail) always pass.  Only extremely blurry or
-    # artificially downscaled images will fail.
-    BLUR_THRESHOLD = 20.0
+    # Threshold lowered drastically because smartphone images often exhibit close-up focus blur
+    # or compression artifacts which lower variance significantly.
+    BLUR_THRESHOLD = 5.0
 
     if lap_var < BLUR_THRESHOLD:
         return {