Research On Pedestrian Attribute Recognition Based On Deep Learning

With the rapid development of the Internet,the Internet of Things,which consists of cameras and other devices,has quietly developed.In some public places such as shops and airports,monitoring equipment fulfills their duties in all corners,and rationally utilizes these monitoring devices to help solve problems such as criminal cases has become an important research topic.In most criminal investigation cases,the characteristics of criminal suspects need to be analyzed,which leads to the research content of pedestrian attribute recognition.Pedestrian attributes are semantic features that can be directly understood by humans.However,for machines,the mapping relationship between pedestrian attribute characteristics and low-level features of images is very complicated.Traditional methods based on low-level feature analysis can only be limited to one of pedestrian attributes.It is difficult to fully exploit the connection between multiple attributes.A pedestrian attribute recognition method based on deep learning is proposed In this paper.This method combines the hidden connections between multiple attributes to realize multi-tag identification of pedestrian attributes.A deep residual multi-branch classification network that introduces attention mechanism is proposed in this paper.The part of the network that extracts image features uses an improved structure based on the ResNet model.Compared with the traditional ResNet model,the order of each layer in the residual network branch is improved,and the feature map outputted in the residual network branch is re-pressed by channel weight,with the weight of the channel reflecting the importance of the channel signature for the final result of the classification.The classification layer of the network is a multi-branch classification network.Each attribute has its own classifier,and the output of the network is the combined result of multiple classifiers.The deep residual multi-branch classification network introduced in this paper has been verified on the PETA database.The model is based on the PyTorch development model,and the operating environment is ubuntu 16.04.The model uses 11400 images for training and 7600 images as test samples.The average recognition accuracy of all tags in the trained model is 0.844.The accuracy ba sed on instance is 0.793.The accuracy rate is 0.879.The recall rate is 0.857,and the F1 score is 0.868.The network attribute recognition effect is tested with the actual photos.The same attributes in the photos get the same feedback,and the different attributes in the photos can be distinguished by the classification network.