Update app.py

app.py

@@ -6,7 +6,6 @@ from transformers import pipeline
 from PIL import Image
 import torch
 import numpy as np
-import cv2
 
 # ======================================
 # LOAD REAL AI MODEL (CPU SAFE)
@@ -48,7 +47,7 @@ def analyze(image):
     img = Image.fromarray(image)
     results = detector(img)
 
-    # Extract real and fake scores explicitly
+    # Extract real & fake scores explicitly
     real_score = 0
     fake_score = 0
 
@@ -60,28 +59,37 @@ def analyze(image):
 
     confidence = round(max(real_score, fake_score), 2)
 
-    # =============================
-    # Smoothness Heuristic
-    # =============================
-    cv_img = cv2.cvtColor(np.array(img), cv2.COLOR_RGB2GRAY)
-    laplacian_var = cv2.Laplacian(cv_img, cv2.CV_64F).var()
+    # ======================================
+    # Multi-Signal Forensic Scoring
+    # ======================================
 
-    # Low texture variance = more AI-like
-    smooth_score = max(0, 100 - laplacian_var)
+    gray = np.mean(np.array(img), axis=2)
 
-    # =============================
-    # Combined Risk Scoring
-    # =============================
+    # Texture / Smoothness detection
+    grad_x, grad_y = np.gradient(gray)
+    laplacian_like = np.abs(grad_x) + np.abs(grad_y)
+    texture_variance = np.var(laplacian_like)
 
-    fake_probability = fake_score
+    smooth_score = max(0, 150 - texture_variance)
+
+    # Facial symmetry heuristic
+    width = gray.shape[1]
+    left = gray[:, :width//2]
+    right = np.fliplr(gray[:, width//2:])
 
-    combined_risk = fake_probability + (smooth_score * 0.3)
+    min_width = min(left.shape[1], right.shape[1])
+    symmetry_diff = np.mean(np.abs(left[:, :min_width] - right[:, :min_width]))
+    symmetry_score = max(0, 100 - symmetry_diff)
+
+    # Combined forensic index
+    fake_probability = fake_score
+    combined_risk = fake_probability + (smooth_score * 0.4) + (symmetry_score * 0.3)
 
-    if combined_risk > 70:
+    if combined_risk > 120:
         prediction = "DEEPFAKE"
         risk_level = "Forensic Risk Level: High"
         risk_color = "#FF3C5F"
-    elif combined_risk > 40:
+    elif combined_risk > 80:
         prediction = "DEEPFAKE"
         risk_level = "Forensic Risk Level: Medium"
         risk_color = "#FFA500"
@@ -94,8 +102,9 @@ def analyze(image):
 • Vision Transformer inference complete
 • Real probability: {round(real_score,2)}%
 • Fake probability: {round(fake_score,2)}%
-• Texture variance score: {round(laplacian_var,2)}
-• Combined forensic risk index: {round(combined_risk,2)}
+• Texture variance score: {round(texture_variance,2)}
+• Symmetry anomaly score: {round(symmetry_score,2)}
+• Combined forensic index: {round(combined_risk,2)}
 • Multi-signal artifact evaluation enabled
 Conclusion: Image classified as {prediction}.
 """
@@ -269,7 +278,7 @@ def get_recent_records():
     return html
 
 # ======================================
-# UI
+# UI (UNCHANGED)
 # ======================================
 
 with gr.Blocks() as demo: