app.py

import gradio as gr
import cv2
import librosa
import numpy as np
from keras.models import load_model
import os
import cv2
import json
import pickle
import librosa
import shutil
import numpy as np
import pandas as pd
from pathlib import Path
from scipy.io import wavfile
from moviepy.editor import VideoFileClip
from keras.utils import np_utils
from sklearn.preprocessing import LabelEncoder

# 데이터 전처리
def preprocess_video(video_path):
    face_cascade = cv2.CascadeClassifier('/content/drive/Shareddrives/23 인공지능 모델링_돌핀/haarcascade_frontalface_default.xml')
    cnn_data = []
    rnn_data = []

    cap = cv2.VideoCapture(video_path)
    count = 0
    while len(cnn_data) < 2:
        ret, frame = cap.read()
        if ret:
            gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
            faces = face_cascade.detectMultiScale(gray, scaleFactor=1.1, minNeighbors=5)
            for (x, y, w, h) in faces:
                face_img = gray[y:y+h, x:x+w]
                resized_img = cv2.resize(face_img, (224, 224))
                cnn_data.append(resized_img)
                count += 1
                if count >= 15:
                    break
        else:
            break

    if len(cnn_data) < 281:
        video_clip = VideoFileClip(video_path)
        audio_clip = video_clip.audio
        audio_clip.write_audiofile("audio.wav")
        y, sr = librosa.load("audio.wav", sr=44100)
        mfcc = librosa.feature.mfcc(y=y, sr=sr, n_mfcc=20)
        mfcc = mfcc[:, :400]
        rnn_data.append(mfcc)
        os.remove("audio.wav")

    cnn_data = np.array(cnn_data)
    rnn_data = np.array(rnn_data)
    return cnn_data, rnn_data

# 딥페이크 영상 유무 판별
def detect_deepfake(video_path):
    cnn_data, rnn_data = preprocess_video(video_path)

    cnn_data_np = np.array(cnn_data)
    rnn_data_np= np.array(rnn_data)

    def augment_data(data, target_size):
      # 증강된 데이터 배열 초기화
      augmented_data = np.empty((target_size,) + data.shape[1:])

      # RNN 데이터를 반전하여 복사
      for i in range(target_size):
          augmented_data[i] = np.flip(data[i % data.shape[0]], axis=0)

      return augmented_data

    # RNN 데이터 증강
    augmented_rnn_data = augment_data(rnn_data_np, cnn_data_np.shape[0])

    y_pred = multimodal_model.predict([cnn_data, augmented_rnn_data])

    #print(y_pred)
    max_prob = np.max(y_pred)
    print(max_prob)

    if max_prob < 0.5:
        result =  "Deepfake"
    else:
        result =  "Real"
    return result
iface = gr.Interface(
    fn=detect_deepfake,
    inputs="video",
    outputs="text",
    title="Video Deepfake Detection",
    description="Upload a video to check if it contains deepfake content.",
    allow_flagging=False,
    analytics_enabled=False
)

iface.launch()