Update main.py

main.py

@@ -94,7 +94,7 @@ async def analyze_video(file: UploadFile = File(...)):
     try:
         # Extract faces
         print("Extracting faces from video...")
-        faces = extract_faces_from_video(temp_video_path, max_frames=6)
+        faces = extract_faces_from_video(temp_video_path, max_frames=15)
         
         if not faces:
             return {
@@ -118,9 +118,15 @@ async def analyze_video(file: UploadFile = File(...)):
         
         # For dima806/deepfake_vs_real_image_detection, Fake is index 1 and Real is index 0
         fake_idx = 1 
+        fake_probs = probabilities[:, fake_idx]
         
-        # Get average probability for 'Fake' class across all analyzed frames
-        avg_fake_prob = probabilities[:, fake_idx].mean().item()
+        # Use Top-K Pooling: Analyze the most "fake" scoring frames
+        # Deepfakes often have glitches in a few frames. Average the worst 3 frames.
+        k = min(3, len(fake_probs))
+        top_k_probs, _ = torch.topk(fake_probs, k)
+        
+        # Get average probability for 'Fake' class across the worst frames
+        avg_fake_prob = top_k_probs.mean().item()
         
         is_fake = avg_fake_prob > 0.5
         
@@ -139,8 +145,9 @@ async def analyze_video(file: UploadFile = File(...)):
             "confidence": confidence,
             "explanation": explanation,
             "details": [
-                {"title": "Face Detection", "desc": f"Analyzed {len(faces)} key frames evenly distributed across the video."},
-                {"title": "Spatial Analysis", "desc": "Evaluated using a Vision Transformer (ViT) deep learning architecture."}
+                {"title": "Face Detection", "desc": f"Extracted {len(faces)} key frames and isolated the subject's face using MTCNN."},
+                {"title": "Spatial Analysis", "desc": "Evaluated using a Vision Transformer (ViT) deep learning architecture."},
+                {"title": "Temporal Pooling", "desc": f"Applied Top-{k} analysis to identify and flag the highest-risk manipulated frames."}
             ]
         }