Ron0420/EfficientNetV2_Deepfakes_Video_Detector

1import gradio as gr2import cv23import numpy as np4import tensorflow as tf5import tensorflow_addons6 7from facenet_pytorch import MTCNN8from PIL import Image9import moviepy.editor as mp10import os11import zipfile12 13local_zip = "FINAL-EFFICIENTNETV2-B0.zip"14zip_ref = zipfile.ZipFile(local_zip, 'r')15zip_ref.extractall('FINAL-EFFICIENTNETV2-B0')16zip_ref.close()17 18# Load face detector19mtcnn = MTCNN(margin=14, keep_all=True, factor=0.7, device='cpu')20 21#Face Detection function, Reference: (Timesler, 2020); Source link: https://www.kaggle.com/timesler/facial-recognition-model-in-pytorch22class DetectionPipeline:23    """Pipeline class for detecting faces in the frames of a video file."""24 25    def __init__(self, detector, n_frames=None, batch_size=60, resize=None):26        """Constructor for DetectionPipeline class.27 28        Keyword Arguments:29            n_frames {int} -- Total number of frames to load. These will be evenly spaced30                throughout the video. If not specified (i.e., None), all frames will be loaded.31                (default: {None})32            batch_size {int} -- Batch size to use with MTCNN face detector. (default: {32})33            resize {float} -- Fraction by which to resize frames from original prior to face34                detection. A value less than 1 results in downsampling and a value greater than35                1 result in upsampling. (default: {None})36        """37        self.detector = detector38        self.n_frames = n_frames39        self.batch_size = batch_size40        self.resize = resize41 42    def __call__(self, filename):43        """Load frames from an MP4 video and detect faces.44 45        Arguments:46            filename {str} -- Path to video.47        """48        # Create video reader and find length49        v_cap = cv2.VideoCapture(filename)50        v_len = int(v_cap.get(cv2.CAP_PROP_FRAME_COUNT))51 52        # Pick 'n_frames' evenly spaced frames to sample53        if self.n_frames is None:54            sample = np.arange(0, v_len)55        else:56            sample = np.linspace(0, v_len - 1, self.n_frames).astype(int)57 58        # Loop through frames59        faces = []60        frames = []61        for j in range(v_len):62            success = v_cap.grab()63            if j in sample:64                # Load frame65                success, frame = v_cap.retrieve()66                if not success:67                    continue68                frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)69                # frame = Image.fromarray(frame)70 71                # Resize frame to desired size72                if self.resize is not None:73                    frame = frame.resize([int(d * self.resize) for d in frame.size])74                frames.append(frame)75 76                # When batch is full, detect faces and reset frame list77                if len(frames) % self.batch_size == 0 or j == sample[-1]:78 79                    boxes, probs = self.detector.detect(frames)80 81                    for i in range(len(frames)):82 83                        if boxes[i] is None:84                            faces.append(face2)     #append previous face frame if no face is detected85                            continue86 87                        box = boxes[i][0].astype(int)88                        frame = frames[i]89                        face = frame[box[1]:box[3], box[0]:box[2]]90 91                        if not face.any():92                            faces.append(face2)     #append previous face frame if no face is detected93                            continue94 95                        face2 = cv2.resize(face, (224, 224))96 97                        faces.append(face2)98 99                    frames = []100 101        v_cap.release()102 103        return faces104 105 106detection_pipeline = DetectionPipeline(detector=mtcnn,n_frames=20, batch_size=60)107 108model = tf.keras.models.load_model("FINAL-EFFICIENTNETV2-B0")109 110 111def deepfakespredict(input_video):112 113    faces = detection_pipeline(input_video)114 115    total = 0116    real = 0117    fake = 0118 119    for face in faces:120 121        face2 = face/255122        pred = model.predict(np.expand_dims(face2, axis=0))[0]123        total+=1124 125        pred2 = pred[1]126 127        if pred2 > 0.5:128          fake+=1129        else:130          real+=1131 132    fake_ratio = fake/total133 134    text =""135    text2 = "Deepfakes Confidence: " + str(fake_ratio*100) + "%"136 137    if fake_ratio >= 0.5:138        text = "The video is FAKE."139    else:140        text = "The video is REAL."141 142    face_frames = []143    144    for face in faces:145        face_frame = Image.fromarray(face.astype('uint8'), 'RGB')146        face_frames.append(face_frame)147        148    face_frames[0].save('results.gif', save_all=True, append_images=face_frames[1:], duration = 250, loop = 100 )149    clip = mp.VideoFileClip("results.gif")150    clip.write_videofile("video.mp4")151 152    return text, text2, "video.mp4"153 154 155 156title="EfficientNetV2 Deepfakes Video Detector"157description="This is a demo implementation of EfficientNetV2 Deepfakes Image Detector by using frame-by-frame detection. \158            To use it, simply upload your video, or click one of the examples to load them.\159            This demo and model represent the Final Year Project titled \"Achieving Face Swapped Deepfakes Detection Using EfficientNetV2\" by a CS undergraduate Lee Sheng Yeh. \160            The examples were extracted from Celeb-DF(V2)(Li et al, 2020) and FaceForensics++(Rossler et al., 2019). Full reference details is available in \"references.txt.\" \161            The examples are used under fair use to demo the working of the model only. If any copyright is infringed, please contact the researcher via this email: tp054565@mail.apu.edu.my.\162            "163            164examples = [              165                ['Video1-fake-1-ff.mp4'],166                ['Video6-real-1-ff.mp4'],167                ['Video3-fake-3-ff.mp4'],168                ['Video8-real-3-ff.mp4'],169                ['real-1.mp4'],170                ['fake-1.mp4'],171           ]172           173gr.Interface(deepfakespredict,174                     inputs = ["video"],175                     outputs=["text","text", gr.outputs.Video(label="Detected face sequence")],176                     title=title,177                     description=description,178                     examples=examples179                     ).launch()