Research On Person Re-identification Based On Multi-scale Attention Fusion

Person Re-identification(Re-ID)also known as pedestrian re-recognition,is a technology that uses computer vision technology to retrieve whether there is a specific pedestrian under cross-device.Pedestrian re-identification technology has broad application prospects in public safety,security monitoring,retail industry and smart cities.However,the current methods have the following problems: First,the existing methods extract some features that interfere with pedestrian recognition due to uncontrollable factors.such as object occlusion,illumination changes,and angle changes;second,most existing methods mainly focus on space attention or channel attention ignore the complementarity between attentions,which limits the model to extract discriminative features,resulting in low accuracy and insufficient robustness of the algorithms.Therefore,this thesis focuses on the application of multi-scale attention fusion in pedestrian re-identification.The main work is as follows:1.This thesis proposes an ECA+SE Attention fusion network(ECA+SE AF-CNN)to solve the problem that most existing methods are not flexible enough when using spatial attention and channel attention in combination.The ECA+SE attention fusion block is used to weight fusion channel attention and spatial attention for improving the attention deficit of a single attention.Specifically,the two kinds of attention can make full use of their own advantages and reduce other disturbances from the channel and spatial attention levels,thus enhancing the feature extraction.The experimental results show that the method can effectively fuse two different attentions,and at the same time has a high recognition accuracy.2.Multi-scale features have the advantage of expanding the receptive field and easily capturing small differences between different feature maps.Therefore,based on the aforementioned research,this thesis further proposes a multi-scale attention fusion network based on ECA+SE.ECA+SE Multi-scale Attention Fusion CNN(ECA+SE MAF-CNN)is used to solve the problem of subtle feature difference loss caused by the existing methods ignoring the receptive fields of different sizes.First,the ECA+SE Multiscale Attention Fusion block is designed to expand the receptive field,capture the small differences between different feature maps,and adaptively focus on the key features and sub-critical features at different scales.Secondly,the ECA+SE AF block is used to extract features from spatial attention and channel attention enhancement.Then,the batch feature discarding method is used to force the network to learn more discriminative features in the local branches,and then the different weights of the generalized average pooling method are used to dynamically enhance the features in the key areas.Finally,a multibranch network is established,which takes into account multi-scale and multi-level feature extraction for obtaining more discriminative fusion features.Experimental results show that the ECA+SE MAF-CNN can effectively take into account the characteristics of pedestrians of different scales,use different size receptive fields to capture more subtle feature differences,and the recognition accuracy rate is also improved.