Research On Person Re-identification Technology Based On Deep Learning

In the area of public safety,person re-identification plays a major role.However,Due to the presence of influencing factors,such as light changes brought on by changes in the monitoring environment,the robustness of the image features extracted by the model is weak,and the detail is insufficient,which increases the difficulty of person re-identification techniques.This work suggests three models with deep learning as the foundation of the research in order to increase the accuracy and boost the feature extraction power of the models.The key components of the research are as follows:(1)A network model based on adversarial consistency and multi-patch characteristics is proposed for the issues of light transformation,various resolutions,and pose changes brought on by variations in the filming equipment in the surveillance environment.In the adversarial consistency learning phase,the dataset is randomly divided into two groups by camera viewpoint,and different feature classifiers are trained for each of the two groups of pedestrian images using cross-entropy and adversarial consistency loss,giving the model the ability to extract robust features that are independent of the filming device.In the multi-patch feature learning stage,the full sample dataset is trained using triplet and multi-patch correlation loss,such that the same pedestrian features in the high-dimensional space are clustered with each other and different pedestrian features are kept away from each other.The experiments show that the proposed model has a good ability to extract effective features.(2)To address the issue of insufficient mining of image detail features caused by the influence of background clutter information in person re-identification algorithms,a network model with a joint normalization module and multi-branch features is developed.In order to enable the network model to focus adaptively on more discriminative features in the image,an attention mechanism-guided instance normalization module is incorporated into the backbone network.The two-level feature fusion module is trained to weight local features before aggregation and fuse them with global features to create a more thorough and detailed feature representation of pedestrian features.To update the network parameters during the model optimization process,joint smoothed cross-entropy loss,triplet loss,and cross-branch feature distillation loss are proposed.The results show that the proposed model enhances the extraction of detailed image features.(3)A joint attention mechanism and a network model with multi-branch features are proposed from the level of alignment of multi-branch features in the maximum difference subspace in order to further lessen the impact of external factors on feature extraction and improve the recognition rate of the person re-identification model.To improve the model’s ability to extract useful characteristics from images,a self-attention mechanism module is integrated into the residual network.After the feature aligner,the global,local,and random erasure features are combined to create a more thorough description of the pedestrian in the deep feature mining module.The model is updated in the optimization process jointly with cosine cross-entropy loss,full-sample triplet loss,center loss and feature alignment loss using a min-max strategy.The effectiveness of the proposed model is verified through comparison and ablation experiments.