Research On Pedestrian Attribute Recognition Based On Multi-task Neural Network

Pedestrian attributes provide rich semantic information of individual humans for pedestrian targets.Pedestrian attribute recognition is a research hotspot in the field of computer vision.It is widely used in the fields of image retrieval,smart security and human-computer interaction,and has a wide range of academic research and practical applications.value.In recent years,convolutional neural networks have made remarkable achievements in the field of visual image recognition.The research on attribute learning based on the deep convolutional neural network framework is constantly progressing,but pedestrian attributes are susceptible to changes in perspective,the diversity of scenes and pedestrian appearances,and differences in attributes within classes.How to improve the generalization and accuracy of the algorithm model,the effect of attribute recognition still needs further research.This paper studies the pedestrian attribute recognition method based on multi-task neural network.First,in the single attribute recognition and multi-attribute recognition tasks of pedestrians,the deep residual network is used for feature extraction and recognition.The results show that there are potential internal connections between multiple pedestrian attributes,joint training and recognition of multiple attributes achieves better results than single-attribute recognition.However,the accuracy of finegrained attribute recognition of pedestrians has not been significantly improved.The analysis reason is that each pedestrian has different attributes,resulting in unbalanced distribution of attribute samples and the influence of non-ideal conditions such as attribute differences within the class.This paper further proposes a multi-task neural network based on the attention mechanism and deep residual network.The attention mechanism is introduced on the basis of the deep residual network,and the residual network combined with the attention mechanism is designed to focus on learning and extracting salient features information;In the network output loss layer,for the imbalance between the positive and negative ratios of the sample,the loss function of dynamic weight is adopted to avoid the forced learning of different attribute tasks with the same weight,and accelerate the model convergence.In this paper,the experiment performed performance verification on the public data set of PETA pedestrian attributes,and set up ablation comparison experimental verification,and analyzed the influence of the improved network structure above on the model recognition performance.The experimental results show that(1)the residual network combined with the attention mechanism is more learning about the related features of different attributes,and focuses on useful feature information to suppress redundant information.(2)The loss function introduces loss weights,which avoids the forced learning of different attribute tasks with the same weight,can adjust different attribute learning tasks,and optimizes the direction of network training.The average recognition accuracy rate of all tags in the improved model in this paper is 0.842,the sample-based accuracy rate is 0.796,the precision rate is 0.873,the recall rate is 0.861,and the F1 score is 0.867.Finally,the actual photos are used to test the recognition effect of the network.The same attributes in pedestrians get the same prediction results,and different attribute networks can be distinguished,which confirms the feasibility of the method in this paper.