A joint improved DETR network face recognition algorithm based on TSM module and DANN.

As one of the most mature application fields of artificial intelligence, face recognition system has been widely used in production and life. But at the same time as large-scale commercialization, face recognition technology also faces more challenges. How to further improve the accuracy of face recognition function and improve the defense ability of the recognition system against face against samples is an important research direction of algorithms. At present, the algorithm mostly focuses on the picture features of a single portrait, ignores the details difference in the time domain of the fake video, and has the problems of weak generalization ability and overfitting of the model. In this paper, an improved DETR network is proposed, which uses the TSM module to perform time domain displacement on the extracted video features and averages the time features of pooled learning samples to better distinguish dynamic adversarial sample instances. At the same time, DANN is introduced as a systematic classification and discrimination network, which uses the domain classifier to domain discrimination of the feature space and uses the adversarial loss function to update the feature extractor and domain classifier parameters. Actual tests show that the recognition accuracy of the network on the FaceForensics dataset is improved by an average of 4%-7% compared with the improvement period, the error rate is less than 7%, and the model recognition speed index is not higher 400ms, which proves that the model has high accuracy rate and good real-time solving ability.