Update app.py

app.py

@@ -7,6 +7,7 @@ import re
 from ultralytics import YOLO
 import easyocr
 from transformers import TrOCRProcessor, VisionEncoderDecoderModel
+from scipy.signal import medfilt
 
 # Load models
 device = "cuda" if torch.cuda.is_available() else "cpu"
@@ -19,12 +20,18 @@ ocr_reader = easyocr.Reader(["en"])  # EasyOCR
 trocr_processor = TrOCRProcessor.from_pretrained("microsoft/trocr-base-stage1")
 trocr_model = VisionEncoderDecoderModel.from_pretrained("microsoft/trocr-base-stage1").to(device)
 
-# Image Preprocessing (Sharpen & Threshold)
+# Image Preprocessing (Sharpen & Adaptive Threshold)
 def enhance_image(image):
     image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
+    
+    # Apply sharpening
     kernel = np.array([[0, -1, 0], [-1, 5, -1], [0, -1, 0]])
     sharpened = cv2.filter2D(image, -1, kernel)
-    _, thresholded = cv2.threshold(sharpened, 150, 255, cv2.THRESH_BINARY)
+    
+    # Apply adaptive thresholding for digital meter reading
+    thresholded = cv2.adaptiveThreshold(sharpened, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C,
+                                        cv2.THRESH_BINARY, 11, 2)
+    
     return thresholded
 
 # Convert Grayscale to RGB (Fix for TrOCR)
@@ -41,7 +48,7 @@ def detect_meter(image):
     for result in results:
         if hasattr(result, "boxes"):  # Ensure correct format
             for box in result.boxes:
-                if box.conf > 0.5:  # Confidence threshold
+                if box.conf > 0.3:  # Lower confidence threshold for better detection
                     detected_meters.append(box.xyxy.tolist())
 
     return detected_meters
@@ -65,6 +72,14 @@ def extract_weight(text):
     matches = re.findall(r'\d+\.\d+|\d+', text)  # Extract numeric weight
     return matches[0] if matches else "Weight not detected"
 
+# Apply Statistical Filtering (Median Filtering for Stable Readings)
+def filter_weight_values(weights):
+    if len(weights) > 1:
+        weights = [float(w) for w in weights]
+        filtered_weight = medfilt(weights, kernel_size=3)[-1]  # Smooth out variations
+        return str(round(filtered_weight, 2))
+    return weights[0] if weights else "Weight not detected"
+
 # Full Processing Pipeline
 def process_image(image):
     enhanced = enhance_image(image)
@@ -76,7 +91,9 @@ def process_image(image):
     weight_easyocr = extract_weight(text_easyocr)
     weight_trocr = extract_weight(text_trocr)
 
-    final_weight = weight_easyocr if weight_easyocr != "Weight not detected" else weight_trocr
+    final_weights = [weight_easyocr, weight_trocr]
+    final_weight = filter_weight_values([w for w in final_weights if w != "Weight not detected"])
+    
     return final_weight or "Weight not detected"
 
 # Gradio Interface