Image-Verifier / app.py
Zeyadd-Mostaffa's picture
Update app.py
e73e2a2 verified
raw
history blame
3.13 kB
import os
import cv2
import numpy as np
import gradio as gr
import tensorflow as tf
from tensorflow.keras.models import load_model
from tensorflow.keras.applications.xception import preprocess_input as xcp_pre
from tensorflow.keras.applications.efficientnet import preprocess_input as eff_pre
from huggingface_hub import hf_hub_download
from mtcnn import MTCNN
import matplotlib.pyplot as plt
# Download models
xcp_path = hf_hub_download(repo_id="Zeyadd-Mostaffa/deepfake-image-detector", filename="xception_model.h5")
eff_path = hf_hub_download(repo_id="Zeyadd-Mostaffa/deepfake-image-detector", filename="efficientnet_model.h5")
# Load models
xcp_model = load_model(xcp_path)
eff_model = load_model(eff_path)
# Face detector
detector = MTCNN()
def extract_face(image):
faces = detector.detect_faces(image)
if not faces:
return None
x, y, w, h = faces[0]['box']
x, y = max(0, x), max(0, y)
return image[y:y+h, x:x+w]
def grad_cam(model, image, size, preprocess_fn):
img = cv2.resize(image, size)
input_tensor = preprocess_fn(img.astype(np.float32))[np.newaxis, ...]
grad_model = tf.keras.models.Model(
[model.inputs], [model.get_layer(index=-1).output, model.output]
)
with tf.GradientTape() as tape:
conv_outputs, predictions = grad_model(input_tensor)
loss = predictions[:, 0] # Assuming binary classification
grads = tape.gradient(loss, conv_outputs)[0]
conv_outputs = conv_outputs[0]
weights = tf.reduce_mean(grads, axis=(0, 1))
cam = np.zeros(conv_outputs.shape[:2], dtype=np.float32)
for i, w in enumerate(weights):
cam += w * conv_outputs[:, :, i]
cam = np.maximum(cam, 0)
cam = cam / (cam.max() + 1e-8)
heatmap = cv2.resize(cam, size)
heatmap = np.uint8(255 * heatmap)
heatmap = cv2.applyColorMap(heatmap, cv2.COLORMAP_JET)
superimposed_img = cv2.addWeighted(cv2.resize(img, size), 0.6, heatmap, 0.4, 0)
return superimposed_img
def predict(image):
face = extract_face(image)
if face is None:
return "No face detected", None
# Xception
xcp_img = cv2.resize(face, (299, 299))
xcp_tensor = xcp_pre(xcp_img.astype(np.float32))[np.newaxis, ...]
xcp_pred = xcp_model.predict(xcp_tensor)[0][0]
# EfficientNet
eff_img = cv2.resize(face, (224, 224))
eff_tensor = eff_pre(eff_img.astype(np.float32))[np.newaxis, ...]
eff_pred = eff_model.predict(eff_tensor)[0][0]
# Ensemble average
avg_pred = (xcp_pred + eff_pred) / 2
label = "Fake" if avg_pred > 0.5 else "Real"
# Grad-CAM on Xception
cam_img = grad_cam(xcp_model, face, (299, 299), xcp_pre)
return label, cam_img
interface = gr.Interface(
fn=predict,
inputs=gr.Image(type="numpy", label="image"),
outputs=[
gr.Label(label="Prediction"),
gr.Image(type="numpy", label="Grad-CAM")
],
title="Deepfake Image Detector (with Grad-CAM)",
description="Upload an image. We detect the face, classify using an ensemble (Xception + EfficientNetB4), and explain the prediction using Grad-CAM on Xception."
)
interface.launch()