Research On Face Recognition Algorithm Based On Deep Learning

Face recognition is a popular and long-standing subject in the field of computer vision.With the gradual deepening of the understanding of deep learning,the convenient acquisition of large-scale data sets and the breakthrough of hardware acceleration platform,deep face recognition technology has also ushered in a qualitative leap and has been widely used in reality.However,for the problems in different applications and scenes in reality,more accurate and stable face recognition and more efficient real-time recognition speed are still a major challenge.This paper mainly improves face recognition from the following three aspects: lighter face detection,more effective feature extraction network and more discriminative loss function.Through the optimization and improvement of these three aspects,we can realize faster and more accurate face recognition.1)In an unconstrained environment,face detection and alignment are always deployed on mobile devices with low computing power and limited storage space,which can not achieve real-time results.A new single-stage lightweight face detection method is proposed.Through the lightweight design of face detection backbone network,real-time face detection under non GPU embedded devices is realized.By introducing a variety of optimization strategies,the detection accuracy can be effectively improved.Compared with several real-time face detection algorithms,this method has obvious advantages in terms of speed and accuracy.2)Aiming at the problem that the existing face recognition network model only aims at feature extraction and ignores feature design,a double attention mechanism network structure is proposed.Firstly,low-level features are designed,detail attention mechanism is used to adaptively learn hierarchical features,multi-scale information is captured through feature pyramid,and attention mechanism is used to encourage the model to pay attention to local features.Secondly,for the design of high-level features,the adaptive semantic grouping of high-level features is carried out through the semantic attention mechanism,and the semantic features are obtained according to the linear weighting between the grouping probability and the corresponding grouping.Finally,the two features are combined to produce high-quality and more discriminative face features.3)In view of the problem that the existing training loss does not consider the marginal itself,resulting in the model’s unclear classification effect on unbalanced samples and difficult convergence in the classification process,an adaptive circular marginal loss adacmloss is proposed,which mainly analyzes the marginal itself and adaptively learns different margins according to different sample categories,so that each sample class has a unique margin.Specifically,By learning more margin from a small number of unbalanced samples,the samples are compressed more compactly in the class of feature space,and the distance between classes is potentially increased while compressing the class,so that the model has stronger generalization ability and clear convergence state.