Research On Human Attributes Analysis Method In Natural Scene

The purpose of human attribute analysis is to obtain the detailed attribute description information of human in the image,including age,gender,clothing,carrying-things,which can verify the identity of human.Human attribute analysis can not only be widely used in security system,effectively extract useful information from massive image and video data to identity verification and search,but also play an important role in many other fields such as pedestrian re-identification.In the task of human attribute analysis in natural scenes,the inaccuracy of human bounding box localization may cause background interference or loss of main feature information for attribute recognition,which will reduce the accuracy of attribute recognition.Moreover,attribute recognition also has many difficulty,like too many kinds of attribute categories,imbalance of the positive and negative samples,and the distribution of hard samles and easy samples,which are both difficulties and entry points.Therefore,this dissertation does a series research about human attribute analysis in natural scenes with consideration of the above problems:1.Object detection method for human attribute recognition.Firstly,based on the importance of target location accuracy to human attribute recognition,the factors that affect the accuracy of object localization are deeply analyzed.Considering the limitations of traditional bounding box regression,this paper introduced a human detection method that based on self attention and bucket mechanism.Then during non-maximum suppression,the measurement of localization accuracy is introduced into soft-NMS to retain high-quality bounding box.And proposing a new multi-task learning loss function to balance classification and localization.All of these work is to achieve high accuracy localization of human,which lays the foundation for human attribute recognition.2.Human attribute recognition method based on attention mechanism and group convolution.Firstly,in order to improve the expression ability and feature extraction ability of the neural network model,the group attention module is embedded into the existing network structure,so as to improve the recognition precision of human attributes.Then,considering the imbalance of the positive and negative samples,and the distribution of hard samles and easy samples in human attributes recognition dataset,an asymmetric multi-label loss function is used to retain the positive samples’ contribution and suppress the easy negative samples.3.Human attribute recognition method based on graph convolution.Firstly,this paper analyzes the problem of human attribute recognition based on multi-label learning deeply.Considering the complex dependency between attributes,a graph convolution module is added to mine and capture the correlation between human attribute labels,and then constructe a multi attribute recognition network based on graph convolution network.