| With the rapid popularization of artificial intelligence and digital image technology,the efficiency of facial recognition has been ahead of other biometric technologies.Especially in the context of epidemic prevention and control,facial recognition has become the most important recognition technology in contactless environments,but in a multi-angle environment,its accuracy fluctuates by the environment,resulting in a decline in recognition accuracy.The identification of multi-angle facial features mainly faces the following challenges:(1)At present,the main problem of multi-angle facial recognition is that the offset of the angle will cause the face to be distorted,and the corresponding feature points cannot be extracted,so that the difference within the class(different angle photos of the same object)is greater than that between classes,which creates a certain bottleneck in the improvement of recognition accuracy.(2)The existing facial database resources with angle offset of the public dataset are not sufficient,the changes of different angles are uneven,and there is no specific angle labeling dataset for training.However,the collective amount of data for angle labeling is too small,which makes it difficult for deep learning model training.In view of the above problems,this paper takes the traditional GAN network as the basic framework,and makes improvements in the network structure to make it more suitable for the identification task of multi-angle facial features,and finally performs feature recognition on the positive image generated by the entire model,and the accuracy results produced show that this method can retain the facial features of different angles to a large extent,and effectively improve the subsequent identification work.In this paper,the following research is carried out:(1)In order to retain the facial feature information of different angles to the greatest extent,the generator is changed from single channel to dual channel,and is responsible for global feature information and local feature information respectively,and finally the feature tensor is fused to form a feature tensor with the same resolution to produce output.(2)For the multi-angle facial recognition task,the loss function in the original framework will be discarded,and a new symmetry loss and identity preservation loss function that conforms to the face is added to improve the output image quality and the robustness of the overall model.(3)In terms of datasets,in view of the insufficient number and limitation of existing datasets,data expansion is carried out,and self-built datasets(to make up for the lack of sufficient angle pairing problems of public datasets)are fused with European and American datasets to make the trained model more generalized.(4)In order to improve the recognition accuracy,an end-to-end mapping module is added to the original recognition network,so that it can improve the low-quality facial image without losing the recognition accuracy of high-quality facial images.Finally,this paper uses self-built and public datasets to test the trained model,and compares and analyzes in multiple aspects,and the experimental results have excellent effects on multiple test standards,which verifies the effectiveness and rationality of the above method. |
