Research On Noise-robust Face Recognition Algorithms Based On Cascaded Convolutional Neural Networks

Face recognition is an important research topic in computer vision,and it has a wide range of real-world applications.For example,face recognition is used to unlock the screen of mobile phones,and identity the faces in intelligent security.In recent years,with the development of deep learning,the accuracy of face recognition has been greatly improved,which surpasses human beings in some scenarios.However,due to the noise,occlusion,illumination and other external factors in real-world applications,the accuracy of face recognition decreases significantly.The captured face images are usually contaminated with noise,which seriously decreases the quality of the images.Therefore,the research on noise-robust face recognition is valuable.The main works of this thesis are summarized as follows:Firstly,in this paper,we propose a cascaded noise-robust deep CNN method for face recognition under noise.In general,there are two ways to deal with the task of face recognition under noise.One way is to take advantage of the image denoising technique,where the input noisy images are denoised,and then these denoised images are fed into the recognition model for identification.However,such a kind of method ignores the relationship between the two tasks,and the operation is tedious.The other way is to directly generate noise-robust face representations.However,due to its simple architecture,some face details may lose.Therefore,in this thesis,we propose a cascaded noise-robust deep CNN method for face recognition under noise.Specifically,a denoising sub-network and a recognition sub-network are jointly trained in a cascaded manner.For the denoising sub-network,dense connectivity is adopted to make full use of shallow and deep features of CNN.Experimental results on public face datasets demonstrate the superior performance of the proposed method over several state-ofthe-art methods.Secondly,in this paper,we propose a noise-robust lightweight face recognition method based on the attention mechanism,which aims to alleviate the problem that the parameters of the traditional network model are too large.Our network consisting of a denoising sub-network and a face recognition sub-network.The channel attention block siguificantly enhances the utilization of informative feature maps and improves the denoising performance.Furthermore,in order to effectively reduce the parameters in the network,we design a lightweight face recognition sub-network based on the depthwise separable convolution.Experimental results on public face datasets demonstrate that our proposed method not only successfully deals with noise,but also has a small amount of model parameters.The proposed method also shows excellent performance over several state-of-the-art methods in terms of recognition accuracy.