Research On Face Recognition Algorithm Under Noise And Occlusion

With the breakthrough of artificial intelligence,human beings have gradually entered the era of intelligence,and bio-identification technology has also widely used.Face recognition is popular as a flexible and convenient technology.Nowadays,face recognition is not only the recognition of personal identity authentication,but also closely related to information such as our property,privacy and other information,so its security requirements are extremely high.The face recognition network based on deep learning achieves 99.55% accuracy in the ideal environment.However,the noise of image transmission in the actual environment reduces the sharpness of the face image,and the clothing such as sunglasses and masks block the important face features,which greatly reduce the recognition accuracy.Therefore,in the face of complex and changeable environment,it is urgent to improve the accuracy of face recognition algorithm and enhance the robustness of the algorithm.This thesis mainly focuses on two kinds of interference,noise and occlusion,combined with lightweight network to improve the performance of algorithm,which can make it better applied in real life.The main research work are as follows:(1)Aiming at the problem of noise interference in real-time face recognition,a lightweight face recognition network based on feature denoising was proposed.Different from the traditional denoising methods,this thesis combined with the theory of deep learning,feature denoising is carried out on the features extracted from the convolution layer to realize the integration of denoising and recognition.The denoising block is designed as the basic block of the network,which can be applied flexibly to each layer of the network.In addition,according to the characteristics of the network,the network is treated with lightened to further optimize the network structure.Under different noise levels,a stable recognition effect is achieved,which improves the robustness of the algorithm well.Compared with other advanced face recognition algorithms,the model is thinner and the recognition accuracy is significantly improved.(2)In order to solve the problem of local occlusion in face recognition,two algorithms are proposed: one is the face image inpainting and recognition algorithm based on the generative adversarial networks(GANs),which uses the GANs characteristics to repair occlusion area image,and combines the proposed lightweight recognition network to extract the complete image features.The other is the face recognition algorithm based on the fusion attention mechanism,which enhances the feature significance through the principle of attention mechanism.Among them,the global attention block focus on the global information,while the spatial attention block perceives the local target features.The algorithm combines the two attention mechanisms to make full use of effective face features in the non-occlusion region,which eliminates occlusion interference and improves the robustness of the recognition algorithm under local occlusion.Experiments show that the proposed two algorithms have effectively improved the recognition accuracy in local occlusion.Among which,the algorithm based on attention mechanism is lighter and more robust.