Update app.py

app.py

@@ -1,91 +1,93 @@
 import gradio as gr
 from transformers import pipeline
 from PIL import Image
-from PIL.ExifTags import TAGS
 import numpy as np
+import cv2
 
-# Model utama deteksi AI
+# Model utama AI-detector
 detector = pipeline("image-classification", model="umm-maybe/AI-image-detector")
-# Model general classifier (backup)
+# Model tambahan general classifier
 general = pipeline("image-classification", model="google/vit-base-patch16-224")
 
-def get_exif_metadata(img):
-    """Cek metadata kamera untuk membantu deteksi foto asli"""
-    try:
-        exif_data = img._getexif()
-        if exif_data is None:
-            return False
-        for tag_id, value in exif_data.items():
-            tag = TAGS.get(tag_id, tag_id)
-            if tag in ["Make", "Model", "Software"]:
-                return True
-        return False
-    except:
-        return False
+def variance_of_laplacian(image):
+    # Menghitung blur menggunakan Laplacian
+    return cv2.Laplacian(image, cv2.CV_64F).var()
+
+def estimate_noise(image):
+    # Menghitung noise sederhana
+    img_gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
+    m = np.mean(img_gray)
+    noise = np.mean((img_gray - m)**2)
+    return noise
 
 def detect_image(img):
     try:
-        # Prediksi dengan AI-detector
-        result1 = detector(img)
-        label1 = result1[0]['label']
-        conf1 = result1[0]['score'] * 100
-
-        # Prediksi dengan model general
-        result2 = general(img)
-        label2 = result2[0]['label']
-        conf2 = result2[0]['score'] * 100
-
-        # Cek metadata kamera
-        has_metadata = get_exif_metadata(img)
-
-        # Weighted hybrid scoring
-        # Skema: AI-detector 50%, Metadata 30%, general model 20%
-        ai_score = conf1 if "artificial" in label1.lower() or "fake" in label1.lower() else 0
-        human_score = conf1 if "human" in label1.lower() or "real" in label1.lower() else 0
-
-        if has_metadata:
-            human_score += 30  # boost jika metadata ada
-        ai_score = min(ai_score, 100)
-        human_score = min(human_score, 100)
-
-        # Normalisasi agar total = 100%
-        total = ai_score + human_score
-        if total == 0:
-            ai_percent = 50
-            human_percent = 50
-        else:
-            ai_percent = round((ai_score / total) * 100, 2)
-            human_percent = round((human_score / total) * 100, 2)
-
-        # Tentukan hasil akhir
-        if ai_percent == 100:
-            final_text = "🖼️ Gambar ini 100% AI"
-        elif human_percent == 100:
-            final_text = "🖼️ Gambar ini asli 100%"
+        # Convert PIL ke OpenCV format
+        img_cv = cv2.cvtColor(np.array(img), cv2.COLOR_RGB2BGR)
+
+        # Blur dan Noise
+        blur_score = variance_of_laplacian(img_cv)
+        noise_score = estimate_noise(img_cv)
+
+        # Metadata kamera
+        metadata_camera = img.info.get("exif") is not None
+
+        # Prediksi AI
+        result_ai = detector(img)
+        label_ai = result_ai[0]['label'].lower()
+        conf_ai = result_ai[0]['score'] * 100
+
+        # Prediksi general
+        result_gen = general(img)
+        label_gen = result_gen[0]['label']
+        conf_gen = result_gen[0]['score'] * 100
+
+        # Hybrid scoring
+        if "artificial" in label_ai or "fake" in label_ai:
+            ai_score = max(conf_ai, 70)
+        elif "human" in label_ai:
+            ai_score = 100 - conf_ai * 0.7
         else:
-            final_text = f"🖼️ Gambar ini {ai_percent}% AI / {human_percent}% Asli"
+            ai_score = 50
+
+        # Tambah bobot metadata kamera
+        if metadata_camera:
+            ai_score -= 10
+
+        # Penalti blur rendah untuk AI photorealistic
+        if blur_score < 100:
+            ai_score += 10  # foto asli biasanya lebih tajam
+        # Penalti noise sangat rendah (AI sering terlalu bersih)
+        if noise_score < 30:
+            ai_score += 5
+
+        # Clamp nilai
+        ai_score = max(0, min(100, ai_score))
+        real_score = 100 - ai_score
 
-        # Tambahkan info model
+        # Output
         output = f"""
 ### Hasil Deteksi:
-{final_text}
+🖼️ Gambar ini {round(ai_score,2)}% AI / {round(real_score,2)}% Asli
 
-**Model AI-detector:** {label1} ({conf1:.2f}%)
-**Model General (ViT):** {label2} ({conf2:.2f}%)
-**Metadata Kamera Ada:** {"Ya" if has_metadata else "Tidak"}
+**Model AI-detector:** {result_ai[0]['label']} ({round(conf_ai,2)}%)  
+**Model General (ViT):** {label_gen} ({round(conf_gen,2)}%)  
+Blur Score: {round(blur_score,2)}  
+Noise Score: {round(noise_score,2)}  
+Metadata Kamera: {'Ada' if metadata_camera else 'Tidak'}
 """
         return output
 
     except Exception as e:
         return f"Terjadi error: {str(e)}"
 
-# UI Gradio
+# Gradio interface
 iface = gr.Interface(
     fn=detect_image,
     inputs=gr.Image(type="pil"),
     outputs="markdown",
-    title="Hybrid AI vs Real Image Detector",
-    description="Upload gambar, sistem akan mendeteksi persentase AI vs foto asli, sensitif untuk AI photorealistic tapi tetap akurat untuk foto asli."
+    title="AI vs Real Image Detector (Advanced Hybrid)",
+    description="Gratis, mendeteksi AI photorealistic dengan blur & noise analisis, semua persentase."
 )
 
 if __name__ == "__main__":