Pedestrian Attribute Recognition Based On Graph And Multi-Scale Network

Pedestrian attributes are usually some external characteristics of pedestrians that can be observed,such as gender,age,type of clothing,and items carried,etc.Given a person image,pedestrian attribute recognition aims at predicting a group of attributes.In the process of acquiring pedestrian data,due to some factors such as imaging equipment and external environment,there are some problems such as low resolution,illumination changes,multi-views,and different poses,which cause serious deformation of attributes and occlusions,resulting in low image quality,an obstacle to subsequent processing and related applications(such as object recognition,pedestrian re-identification,etc.).Traditional pedestrian attributes recognition methods are generally based on the feature of hand-design.In recent years,pedestrian attribute recognition methods are mainly based on deep learning for the convolution neural network's outstanding performance in computer vision.In addition,a pedestrian image generally has multiple attributes with different levels of semantic features and different sizes,and intricate label space.Therefore,some of the fine-grained attributes features are eliminated after some operations(such as convolution and pooling)during deep learning and the relationship between tags cannot be fully utilized.According to the above problems,the main contributions of this dissertation are summarized as follows:1)Pedestrian attribute recognition tasks are analyzed,and the basic knowledge of deep learning technology,multi-label image classification and graph convolution network are introduced;2)A pedestrian attribute recognition algorithm based on graph convolution network and multi-scale network is proposed.The algorithm uses a multi-scale attribute perception module to extract the features of multiple scales which are utilized as the node of the graph convolution module to model the dependence between pedestrian attribute labels;3)A pedestrian attribute recognition algorithm based on graph reasoning is proposed.This algorithm directly performs graph convolution operation on the feature map extracted by CNN,and constructs the adjacency matrix of graph by using self-attention mechanism,which can complete inductive learning task without other prior information;4)Experiments are conducted on two widely used datasets to verify the effectiveness of the proposed algorithms,i.e.the RAP dataset and the PETA dataset.