Update app.py

app.py

@@ -41,7 +41,7 @@ material_params = {
     "mirror": {"alpha": 0.7, "eps": 0.1, "I": 2000, "name": "آینه"},
     "foliage": {"alpha": 0.25, "eps": 0.98, "I": 900, "name": "گیاهان"},
     "water": {"alpha": 0.06, "eps": 0.98, "I": 4200, "name": "آب"},
-    
+    "sky": {"alpha": 1.0, "eps": 1.0, "I": 0, "name": "آسمان"},
 }
 
 material_categories = {
@@ -72,6 +72,10 @@ material_categories = {
     "water_bodies": {
         "members": ["water"],
         "candidates": ["water"]
+    },
+    "background": {
+        "members": ["sky"],
+        "candidates": ["sky"]
     }
 }
 
@@ -104,6 +108,9 @@ replacement_text = {
     },
     "water_bodies": {
         "water": "حفظ منابع آبی به عنوان عنصر خنک‌کننده"
+    },
+    "background": {
+        "sky": "عنصر طبیعی بدون نیاز به تغییر"
     }
 }
 
@@ -133,17 +140,23 @@ def calc_deltaT(material, T_air, RH, u, S):
     gamma = S / max(h_c, 1e-6)
     return gamma * C_m / 1000.0
 
-def get_patches(image, size=224, stride=200):
+def get_patches(image, size=224, stride=100):
     patches = []
     w, h = image.size
     
-    for i in range(0, w, stride):
-        for j in range(0, h, stride):
-            box = (i, j, min(i+size, w), min(j+size, h))
-            patch = image.crop(box)
+    for scale in [1.0, 0.75]:
+        scaled_w, scaled_h = int(w * scale), int(h * scale)
+        if min(scaled_w, scaled_h) < size:
+            continue
             
-            if patch.size[0] >= size//2 and patch.size[1] >= size//2:
-                patches.append(patch)
+        scaled_img = image.resize((scaled_w, scaled_h), Image.Resampling.LANCZOS)
+        
+        for i in range(0, scaled_w, stride):
+            for j in range(0, scaled_h, stride):
+                box = (i, j, min(i+size, scaled_w), min(j+size, scaled_h))
+                patch = scaled_img.crop(box)
+                if patch.size[0] >= size and patch.size[1] >= size:
+                    patches.append(patch)
     
     return patches
 
@@ -158,8 +171,9 @@ def analyze_image(img, T_air, RH, u, S):
         if len(patches) == 0:
             return "⛔ تصویر نامعتبر است یا کوچک است."
         
-        # پیش‌بینی برای تمام پچ‌ها
+        # پیش‌بینی با اعتماد بیشتر
         all_predictions = []
+        confidence_threshold = 0.7  # افزایش آستانه اطمینان
         
         for patch in patches:
             inputs = processor(images=patch, return_tensors="pt")
@@ -167,21 +181,26 @@ def analyze_image(img, T_air, RH, u, S):
             with torch.no_grad():
                 outputs = model(**inputs)
                 probs = torch.nn.functional.softmax(outputs.logits, dim=-1)
+                confidence, pred = torch.max(probs, dim=-1)
             
-            top1 = torch.argmax(probs[0]).item()
-            label = model.config.id2label[top1]
-            all_predictions.append(label)
+            if confidence.item() > confidence_threshold:
+                label = model.config.id2label[pred.item()]
+                all_predictions.append(label)
         
         if not all_predictions:
-            return "⛔ هیچ مصالح معتبری شناسایی نشد."
+            return "⛔ هیچ مصالح معتبری با اطمینان کافی شناسایی نشد."
         
         counter = Counter(all_predictions)
-        total_patches = len(patches)
+        total_patches = len(all_predictions)
         
         # فیلتر مواد با فراوانی کافی
         materials_found = {}
+        ignore_classes = ["food", "skin", "other", "wallpaper", "carpet", "fabric"]
+        
         for m, count in counter.items():
-            if m in material_params and count >= max(3, total_patches * 0.05):
+            if (m in material_params and 
+                m not in ignore_classes and 
+                count >= max(3, total_patches * 0.1)):
                 materials_found[m] = count
         
         if not materials_found: