Research On Pedestrian Attribute Recognition Method Based On Deep Learning

As advanced semantic information,pedestrian attributes can be applied to multiple tasks to assist in improving task performance.The wide range of applications makes pedestrian attribute recognition tasks of great research value.With the rise of deep learning,pedestrian attribute recognition methods based on deep learning have become the mainstream research direction.Under the framework of deep learning technology,this thesis conducts a series of studies that comprehensively consider attribute feature extraction,correlation between attributes,and differences between attributes.The main contents and contributions are as follows:(1)Pedestrian attribute recognition method based on multi-stage feature separation network.In this thesis,the attention mechanism is introduced to build a feature separation module for attributes.Based on the public features initially extracted from the backbone network,the author pays attention to the areas where attribute features exist and extracts attribute features separately for recognition.Aiming at the problem of attention deviation in the case of insufficient supervision signals,a multi-stage feature separation network is designed.By weakening the features of the concerned areas in the shared features of the previous stage,the feature separation module of the network in this stage is forced to take the initiative to pay attention to other areas on the common features,and in the last stage,the attribute features extracted from all the preceding stages are fused and then paid attention to.The most valuable attribute features in the attribute features obtained at each stage are extracted.After attribute recognition at each stage,the result is fused as the final recognition result.Finally,the effectiveness of the proposed method is verified by experiments on public data sets.(2)Pedestrian attribute relationship analysis based on graph convolutional neural network.In this thesis,aiming at the lack of relationship analysis between attributes in the identification process of multi-stage feature separation network,an attribute relationship aggregation module based on the graph convolution network is proposed on the basis of the multi-stage feature separation network.The existing attribute features are taken as attribute nodes,and the attribute relationship is modeled by combining the data-driven and network self-learning methods to build the attribute relationship diagram.The attribute relationship aggregation module is used to carry out the graph convolution operation on the attribute graph,which guides the information propagation among the relevant attribute nodes in the attribute graph and carries out attribute recognition.Other relevant attribute features are introduced as auxiliary information during the network attribute recognition to further improve the network performance.(3)Pedestrian attribute recognition method based on dynamic weighting function.Aiming at the problem that the static weighted loss function,which is commonly used in the field of pedestrian attribute recognition at present,ignores the different degree of training difficulty of samples with different attributes,this thesis transforms it and proposes a dynamic weighted loss function.In the training process,the network is periodically used to propagate forward the training set samples and calculate the loss of each attribute sample.Evaluate the training difficulty of various attribute samples,dynamically adjust the weight corresponding to each attribute in the training loss function,so that the network pays more attention to the difficult training attribute samples in the training process,and thus improve the recognition accuracy of the network on difficult attributes.