The Research And Design Of Face Forensics Algorithm Based On Deep Learning In Transform Domain

With the rapid development of face synthesis technology,the forged face image is becoming more and more realistic,and it is difficult for human eyes to recognize whether the image has been tampered.At the same time,due to the commercialization of tampering technology,the threshold for performing face synthesis has become lower,which poses serious social credibility and security problems.On top of that,images are often processed by compression algorithms such as JPEG during the dissemination process,resulting in artificial artifacts in the images that are more difficult to discern.Therefore,face forgery detection algorithms with high accuracy and robustness are urgently needed.The current mainstream approachs are to combine deep learning with face tampering forensics to extract pixel-level features in the spatial domain for classification,which achieves high accuracy,but the interpretability and robustness are not ideal.The recently emerged algorithms combining transform domain and deep learning have also achieved excellent results and provided new ideas for the interpretability of face forensics.In this thesis,we combine the dual-tree complex wavelet transform and discrete cosine transform with deep learning for common face image tampering methods,and construct two effective algorithmic models for classifying real and fake face images,the main work and innovation of the thesis are as follows:(1)On the basis of the dual-tree complex wavelet transform and deep learning,a face forensics algorithm based on the discordency of face orientation features is proposed.It is found that some artificial artifacts in tampered face images are more obvious in the transform domain than in the spatial domain.Moreover,the geometric structure of the face is relatively stable and there exists a specific distribution pattern,and the artifacts produced by face tampering techniques can destroy this correlation.Therefore,we extract the features of the original image in six directions(±75°,±45°,± 15°)by using the dual-tree complex wavelet transform,and then obtain the correlation between each direction by using the proposed DCE(Directional Correlation Extraction)module,and input into the classification network Resnet to train the final classification model DCWNet.In addition to achieving an average accuracy of 97.3%to 99.6%on general public datasets,the algorithm is also quite robust for compressed images compared with existing algorithms.(2)Based on discrete cosine transform and deep learning techniques,a face forensics algorithm is proposed by enhancing the effective features in the frequency domain.We experimentally illustrate that different frequency components of face images have different contributions to the final discrimination results.Unlike other classes of natural images,the frequency characteristics of face images are relatively stable.This work proposes an adaptive frequency domain enhancement mechanism AFE based on the principle of 2D discrete cosine transform and JPEG compression to enhance the spatial domain features by frequency domain during the model training.This mechanism not only increases the weight of important features,but also reduces the impact of high frequency noise caused by image compression algorithms such as JPEG.In this thesis,we combine AFE with the general classification network Xception and further improve it to propose a face forensics network FENet,which achieves an average accuracy of 98%and high JPEG compression robustness on the public datasets.