Update app.py

app.py

@@ -1,4 +1,3 @@
-
 import cv2 as cv
 import numpy as np
 import gradio as gr
@@ -7,7 +6,7 @@ from huggingface_hub import hf_hub_download
 
 # Download ONNX model from Hugging Face
 model_path = hf_hub_download(repo_id="opencv/image_classification_mobilenet", filename="image_classification_mobilenetv1_2022apr.onnx")
-top_k = 10  # Increased to support variable top_n
+top_k = 10
 backend_id = cv.dnn.DNN_BACKEND_OPENCV
 target_id = cv.dnn.DNN_TARGET_CPU
 
@@ -18,42 +17,58 @@ def add_hsv_noise(image, hue_noise=0, saturation_noise=0, value_noise=0):
     """Add HSV noise to an image"""
     if image is None:
         return None
-
+    
     # Convert BGR to HSV (OpenCV uses BGR by default)
     hsv = cv.cvtColor(image, cv.COLOR_BGR2HSV).astype(np.float32)
-
+    
     # Add noise to each channel
     hsv[:, :, 0] = np.clip(hsv[:, :, 0] + hue_noise, 0, 179)  # Hue: 0-179
     hsv[:, :, 1] = np.clip(hsv[:, :, 1] + saturation_noise, 0, 255)  # Saturation: 0-255
     hsv[:, :, 2] = np.clip(hsv[:, :, 2] + value_noise, 0, 255)  # Value: 0-255
-
+    
     # Convert back to BGR
     bgr = cv.cvtColor(hsv.astype(np.uint8), cv.COLOR_HSV2BGR)
-
+    
     return bgr
 
 def classify_image_with_noise(input_image, top_n, hue_noise, saturation_noise, value_noise):
-    """Classify image with HSV noise applied"""
+    """Classify image with HSV noise applied and return exact confidence scores"""
     if input_image is None:
         return None, "Please upload an image first."
-
+    
     # Apply HSV noise
     noisy_image = add_hsv_noise(input_image, hue_noise, saturation_noise, value_noise)
-
+    
     # Resize and crop as in original code
     image = cv.resize(noisy_image, (256, 256))
     image = image[16:240, 16:240, :]
-
-    # Update model's topK for this inference
-    model.topK = top_n
-    result = model.infer(image)
-
-    # Format results with probabilities if available
-    result_str = "\n".join(f"{i+1}. {label}" for i, label in enumerate(result[:top_n]))
-
+    
+    # Preprocess manually to get raw scores
+    input_blob = model._preprocess(image)
+    
+    # Forward pass
+    model.model.setInput(input_blob, model.input_names)
+    output_blob = model.model.forward(model.output_names)
+    
+    # Get raw probabilities (apply softmax if needed)
+    raw_scores = output_blob[0]  # First batch
+    probabilities = np.exp(raw_scores) / np.sum(np.exp(raw_scores))  # Softmax
+    
+    # Get top N indices and their scores
+    top_indices = np.argsort(probabilities)[::-1][:top_n]
+    
+    # Format results with exact confidence scores
+    result_lines = []
+    for i, idx in enumerate(top_indices):
+        label = model._labels[idx]
+        confidence = probabilities[idx]
+        result_lines.append(f"{i+1}. {label}: {confidence:.6f} ({confidence*100:.4f}%)")
+    
+    result_str = "\n".join(result_lines)
+    
     # Convert BGR to RGB for display in Gradio
     display_image = cv.cvtColor(noisy_image, cv.COLOR_BGR2RGB)
-
+    
     return display_image, result_str
 
 def clear_output_on_change(img):
@@ -73,42 +88,32 @@ with gr.Blocks(css='''.example * {
 
     with gr.Row():
         with gr.Column():
-            # Input controls
             image_input = gr.Image(type="numpy", label="Upload Image")
-
+            
             gr.Markdown("### Classification Settings")
             top_n = gr.Slider(minimum=1, maximum=10, value=5, step=1, label="Top N Classes")
-
+            
             gr.Markdown("### HSV Noise Controls")
             hue_noise = gr.Slider(minimum=-50, maximum=50, value=0, step=1, label="Hue Noise (-50 to 50)")
             saturation_noise = gr.Slider(minimum=-100, maximum=100, value=0, step=5, label="Saturation Noise (-100 to 100)")
             value_noise = gr.Slider(minimum=-100, maximum=100, value=0, step=5, label="Value/Brightness Noise (-100 to 100)")
-
+        
         with gr.Column():
-            # Output displays
             noisy_image_output = gr.Image(label="Image with Noise Applied")
-            output_box = gr.Textbox(label="Top Predictions", lines=10, max_lines=15)
+            output_box = gr.Textbox(label="Top Predictions with Confidence Scores", lines=10, max_lines=15)
 
-    # Clear outputs when new image is uploaded
     image_input.change(fn=clear_output_on_change, inputs=image_input, outputs=[output_box, noisy_image_output])
 
     with gr.Row():
         submit_btn = gr.Button("Submit", variant="primary")
         clear_btn = gr.Button("Clear")
 
-    # Set up real-time updates for sliders
     inputs = [image_input, top_n, hue_noise, saturation_noise, value_noise]
     outputs = [noisy_image_output, output_box]
-
-    # Update predictions when sliders change (real-time)
+    
     for slider in [top_n, hue_noise, saturation_noise, value_noise]:
-        slider.change(
-            fn=classify_image_with_noise,
-            inputs=inputs,
-            outputs=outputs
-        )
-
-    # Manual submit button
+        slider.change(fn=classify_image_with_noise, inputs=inputs, outputs=outputs)
+    
     submit_btn.click(fn=classify_image_with_noise, inputs=inputs, outputs=outputs)
     clear_btn.click(fn=clear_all, outputs=[image_input, noisy_image_output, output_box])
 
@@ -123,4 +128,4 @@ with gr.Blocks(css='''.example * {
     )
 
 if __name__ == "__main__":
-    demo.launch()
+    demo.launch()