Research On Sensorless Face Recognition And Pedestrian Re-Identification Methods With Small Samples

Information technology is developing rapidly,and the pedestrian tracking technology based on face recognition and pedestrian re-identification of machine vision is widely used.However,in most cases,the face recognition task is completed based on the no-sense case,and there is insufficient sample data of face base images,which leads to the low accuracy of no-sense face recognition under small samples;and in the cross-camera case,the re-recognition task has a single model training data and is limited by the pixel scale,which leads to the accuracy of cross-camera pedestrian re-recognition under small samples has been unable to be effectively improved.To solve the above problems,this paper summarizes and analyzes the existing small-sample learning algorithms and proposes the following research contents.1.To address the problem of low accuracy of sensorless face recognition under small samples,a 3DFACE-Cycle GAN face-based multi-angle information feature data enhancement model is built,and the face texture feature extraction capability of the model is enhanced by adding ontology mapping loss to the loss function.The model can generate a large number of rotated multi-angle face image communities from a small number of frontal face images,and generate a base population of rotated multi-angle face image feature information using the fitted feature extraction network.Experiments show that by enriching the base library of face features,the intra-class distance between the face features detected in real time and the target features of the face base library can be reduced,and the inter-class distance between the real-time face features and the non-target features of the face base library can be increased,and the task of sensorless face recognition can be accomplished with high accuracy.2.To address the problem of low accuracy of cross-camera pedestrian re-recognition under small samples,a Cycle GAN-based cross-camera pedestrian image data enhancement model is built,and a residual module and circular convolution structure are added to the model structure to enhance the linear representation of deep features to shallow features.The model can retain more shallow details of the output image and avoid the problem of low accuracy of pedestrian re-recognition due to the large differences in angle,light and background of the same pedestrian image taken by different cameras.The experiments show that by constructing a cross-camera pedestrian image data generation model and data enhancement of cross-camera pedestrian images,the generalization of the re-recognition model can be improved and the cross-camera pedestrian re-recognition task can be accomplished with high accuracy.