Fine-Grained Person Attribute Recognition Based On Multi-Scale Feature Fusion

Pedestrian attribute recognition is the semantic analysis of a given pedestrian image,which plays an important role in video surveillance,identity recognition,person retrieval and other fields.Due to the multiple categories and scales of pedestrian attribute labels,the correlation between attributes is very complex and easy to be ignored,resulting in many features in the image that are not strong in semantic information but play an important role in classification are difficult to be captured by the network.In addition,the size of existing pedestrian attribute datasets is limited by quantity and diversity,with insufficient fine-grained attributes and poor quality.Therefore,the existing methods have poor recognition performance for fine-grained pedestrian attributes.In view of the above problems,this thesis improves and enriches a pedestrian attribute dataset containing 126 attributes for fine-grained pedestrian attribute recognition,and proposes three methods for pedestrian attribute recognition.The main research contents of the thesis are as follows:(1)This thesis analyzes and perfects a fine-grained pedestrian attribute recognition dataset.This dataset covers 126 attributes,including many fine-grained pedestrian attributes,but too many attribute labels lead to uneven data distribution in the dataset,and the problem of long tail is serious,which reduces the quality of the dataset.Aiming at this problem,this thesis improves and expands the data set to enrich the tail data.(2)This thesis studies the pedestrian attribute recognition method based on multiscale feature fusion of attention mechanism.In view of the challenges brought by the diversity of attribute scales,this thesis adopts the form of pyramid to fuse multi-scale features to improve the accuracy of attribute recognition;In order to suppress the interference of irrelevant information and make the model locate the effective information,this thesis introduces the channel attention mechanism to make the network pay attention to more important channel features;In this thesis,the multi-branch loss function is used to constrain,fully train each fused feature layer,and improve the network performance.(3)This thesis studies the pedestrian attribute recognition method based on local feature and multi-granularity feature fusion.There are many small and medium-sized attributes with strong regional correlation in the pedestrian image.Only using global features can not effectively improve the recognition performance of these attributes.To solve this problem,this thesis studies a method combining global and local feature features for attribute recognition,which extracts features from each granular area of pedestrians,effectively improving the recognition accuracy of fine-grained attributes;For the long tail problem in the dataset,this chapter uses the attribute weighted loss function to avoid excessive punishment of tail samples and improve the network recognition performance.(4)In this thesis,a pedestrian attribute recognition method based on graph convolution network is studied.This thesis uses graph to model the interdependencies of attribute labels,so as to flexibly obtain the topological structure of attribute label space and improve the learning ability of the model;In view of the unbalanced distribution of data in the attribute categories of the dataset,the balanced loss distribution loss is adopted to alleviate the imbalance of various categories and improve the performance of the model.