Research Of Generalizable Person Re-Identification Method Based On Deep Features

Person re-identification is one of the research hotspots in the field of image retrieval,which greatly promotes the development and application of intelligent video analysis technology across devices and scenes.In the field of public safety,person re-identification technology can help public security organs quickly screen suspicious individuals and prevent security incidents.In the new retail field,by obtaining customers’ behavior trajectories,it helps businesses obtain customers’ digital information,thereby mining more business value.However,currently,person re-identification faces many challenges.It is affected by factors such as camera angle,acquisition scene,and pedestrian posture.The originally high-performance model will experience a significant decrease in accuracy when transferred to a unseen domain.To address the problem of poor model generalization ability,we conduct research on domain generalization of person re-identification and proposes two algorithms that can effectively improve the generalization,and the main work is as follows.We conducted extensive experiments and found that layer normalization(LN)does not outperform other normalization techniques in transformer-based generalized Re ID.In this thesis,we propose a learnable adaptive normalization approach by combining instance normalization(IN),batch normalization(BN),and layer normalization(LN).First of all,AN is freer to choose more appropriate data for normalization and use IN and BN to reduce the style discrepancy between source and target domains,thus improving the generalization of Transformer on Re ID tasks.Secondly,the AN learns unique normalization parameters for layers of different depths,thus enabling deep and shallow layers to focus on different information.AN can learn unique normalization parameters for layers at different depths,so that deep and shallow layers can pay attention to different information.In addition,this thesis proposes a mixture expert algorithm based on batch normalization,using which the adaptive normalization module is effectively improved to further enhance the generalization performance of the model.Extensive experiments demonstrate that our AN outperforms any existing normalization technique on the transformer-based Re ID model and significantly improves the generalizability of the models.In order to reduce the distribution discrepancy between the source and target domains and thus improve the generalization of the Re ID model,this thesis proposes a DG Re ID algorithm based on distribution alignment.Specifically,this thesis first introduce an intermediate latent space constrained to a known prior distribution.The source domain data is mapped to this latent space and then reconstructed back so that both the source domain distribution and the target domain distribution are aligned to the prior distribution.The similar source domain distribution and target domain distribution bring better generalization to the Re ID model.In addition,we propose a cross-domain self-distillation algorithm module to learn domain invariant features through multiple different domains to further improve the generality of global features.Our approach achieves the best performance under most domain generalizable person re-identification settings.