Research On Face Detection And Occluded Re-ID Algorithm Based On Deep Learning

This work is mainly composed of two parts:face detection and pedestrian reidentification,which respectively correspond to two different application scenarios.When high-quality images of pedestrians are available,they can be judged and identified by face detection and recognition technology;when environmental problems such as occlusion problems and lighting problems are encountered,pedestrian reidentification technology can be used to judge and recognize pedestrians.The main contribution is:1.Face Detection Based on Receptive Field Enlarged Multi-Task Cascaded Convolutional Neural Networks.Aiming at the small receptive field of the classic face detection network Multi-Task Cascaded Convolutional Network(MTCNN),the problem of insufficient discernibility and robustness of small target features is improved.So we proposed Receptive Field Enhanced MTCNN(RFE-MTCNN),by using Inception-V2 module and Receptive Field Block(RFB)module to enlarge the network receptive field.And Global Average Pooling(GAP)is used to replace the fully connected layer,which reduces a large number of parameters brought by the original fully connected layer,and the overall network parameters are reduced by 16%compared to MTCNN.On this basis,the AM-Softmax loss function is introduced into R-Net to enhance the classification performance of R-Net.The experimental results show that compared with MTCNN,the performance of our network on the AFW,PASCAL,and FDDB data sets is improved by 1.08%,2.84%,and 1.31%,respectively,and the accuracy of the three sub-data sets of WIDER FACE is improved by 2.3%,2.1%and 6.6%.2.Occluded Re-ID based on semantic feature matching.In the network training task of pedestrian re-identification,the semantic segmentation task and the pedestrian re-identification task are combined to constrain each other.In addition to the ternary loss function commonly used in pedestrian re-identification tasks,it also adds semantic segmentation loss function.The backbone network and the semantic segmentation network are used to obtain the feature information of each part of the human body,and then perform Global Max Pooling(GMP)to obtain a one-dimensional vector with salient features.Use Gated Recurrent Neural Network(GRU)to encode a onedimensional vector to help the network find potential related information within the one-dimensional feature.Finally,the Convolutional Block Attention Module(CBAM)module is introduced in the different layers of the backbone network ResNet50 to strengthen the extraction of model related features and enhance the expression ability of the model.Experimental results show that Rank-1 in the Market 1501 datasets has increased by 2.1%,and mean Average Precision(mAP)has increased by 4.3%.In the occlusion datasets Partial-REID,Partial-iLIDS 18,Partial-iLIDS 19,Occluded-REID and Occluded-Duke achieved 87%,72.3%,72.9%,87.4%,and 56.7%of Rank-1 accuracy,all of which have reached relatively high accuracy,all of which have reached relatively high accuracy.