Development Of Face Detection And Recognition Algorithm Based On Lightweight Convolution Neural Network

Recently,face recognition technology has gradually become authentication method of access in the banks,financial institutions and government agencies.Although,image processing algorithm based on convolution neural network has efficient,robustness and generalization ability,it has large amount of parameters and calculation.Therefore,the development of fast,lightweight and accurate face detection and recognition algorithm based on convolution neural network has research significance and engineering application value on low-cost,low-power and low-computational processing platform.This thesis researches the face detection and recognition algorithm based on lightweight convolution neural network.The algorithm can accurately detect the faces in a video and identify their identities on ARM Cortex-A72 platform.Based on proposal region regression and object detection algorithm,this thesis analyses the size,scale and shape of faces and applies regression algorithm on multiple feature maps to calculate face detection boxes in order to match multi-size and multi-angle faces in the nature scene.Meanwhile,face detection model utilizes by ways of rapidly down sampling convolution structure,inception-V2 and bottleneck to reduce the computation complexity.In addition,based on convolution neural network,this thesis develops a feature extraction model and uses improved Softmax classifier to train this model.After extracting face feature,it uses multi-classes SVM classifier to finish the 1:N mode face recognition job,and uses cosine similarity model to finish the 1:1 mode face recognition job.Meanwhile,depend on batch normalization this thesis trains feature extraction model with big batch size.Finally,after calculating the formula of convolution and batch normalization,we find a method to merge parameters of these two operators and optimize the amount of calculation and parameters in feature extraction model.Finally this thesis tests face detection algorithm and face recognition algorithm respectively,and combines these two algorithms to satisfy basic functions in face recognition gate system.The algorithm costs about 89ms and achieves more than 99%accuracy in a variety of real scenarios.Moreover,it also adapts to complex background and illumination variation.