基于空洞卷积的低质量人脸深度伪造图片检测

Detection of low⁃quality facial deepfake image based on void convolution

图像分类任务中,在使用卷积神经网络提取图像特征信息时,通过最大池化和下采样的操作提高模型感受野的同时会带来特征图分辨率下降、上下文信息丢失等问题,在原本分辨率较低的人脸深度伪造图像中,局部有效特征信息的丢失会对分类准确度带来较大的影响。基于以上存在的问题,文中提出基于空洞卷积多尺度信息融合的人脸深度伪造检测方法。该方法采用全卷积Resnet34模型和基于空洞卷积的多通道特征融合的结构,利用空洞卷积网络在提高网络感受野的同时不丢失特征图信息的优势,融合多尺度空间特征信息,最大程度地捕捉图像上下文信息。实验在Faceforensic++、Celeb⁃deepfakeforensics、Deepfake in the wild、DFDC previous数据集上均取得比当前其他方法更好的效果。实验结果表明,提取高层语义信息后进行多尺度特征融合,可以提高分辨率低的人脸深度伪造图像的分类效果。

When extracting image feature information by convolutional neural networks in image classification task,the issues of feature resolution reduction and context information loss will be occur while increasing the receptive field of the model by means of the max⁃pooling and down⁃sampling.In the facial deepfake image with low resolution,the loss of local effective feature information will have a great impact on the classification accuracy.Based on the above problems,a void convolution based facial deepfake detection method with the multi⁃scale information fusion is proposed.In this method,the fully convolutional Resnet34 model and the multiple channels feature fusion structure based on void convolution are adopted.While the advantage of void convolutional network is used to improve the receptive field of network without losing the information of feature map,the multi⁃scale spatial feature information is fused to capture the image context information to the greatest extent.The experimental results on the datasets of Faceforensics++,Celeb⁃deepfakeforensics,Deepfake in the wild and DFDC previous are better than other current methods.The results show that the multi⁃scale feature fusion after extracting high⁃level semantic information can improve the classification accuracy of low⁃resolution facial deepfake images.