Person Re-Identification Based On Multiple Granularity Feature Extraction

Person re-identification technology is to study the recognition and retrieval of the one people acrossing cameras and scenes.It is to recognize pedestrians according to their clothing,posture,hair and other information.Current research trends it shows that an important way to improve the performance of pedestrian re-identification tasks is to combine global features with local features.Previously,local-based methods mostly rely on locating regions with specific semantics to learn the representation of local features,which not only increases complicates of the network learning,and it makes the information uncomprehensive.This paper mainly refers to the multi-granularity network and integrates the discriminant information extracted from different granularity under the end-to-end feature learning strategy,one branch is used to extract the global features of pedestrians,and the other two branches are used to extract the local features.Therefore,this paper mainly modified the fully connected part of the end of the network to make it re-identify for pedestrian the combination of features provides more choices and enhances the robustness of the overall model.The loss function of label smoothing is introduced when the model does not approach the predicted labels excessively during the training process to avoid overfitting.This model uses the method of evenly dividing the feature map into horizontal stripes and changing it in different local branches,and vary the number of parts to obtain local feature representations with multiple granularities.In this paper,a comprehensive experiment is performed on the mainstream Market-1501 evaluation data set.The experimental results showed that the multi-granularity network model was replicated.Taking the single query mode as an example through reordering algorithm,Rank-1 and mAP ranked 95.0%and 93.1%respectively.The improved multi-granularity network model can achieve better performance,which has improved 0.8 and 0.6 percentage points on Rank-1 and mAP,respectively,and it is also superior to previous local or local and global feature phase comparisons.