Research On Pedestrian Re-identification Method Based On Local Convolutional Baseline Network And Multi-stage Knowledge Distillation

Person Re-identification(Re ID)refers to the identification of designated person in image sequences or videos taken by multiple non-overlapping cameras.This technology is particularly important for security surveillance,smart retail and other fields.Because the captured person images have uncontrollable factors such as light intensity,background and occlusion,thereby increasing the difficulty of identifying person.The Re ID method based on local features has potential robustness in solving these problems.However,it only focuses on local area information,and ignores complementary with global information,which limits the further improvement of recognition performance.On the other hand,although the task of person reidentification has made remarkable achievements in the field of computer vision,it places high requirements on the amount of calculation and memory,which limits the application of deep neural networks in actual deployment.Knowledge Distillation(KD)provides a promising solution for building lightweight deep learning models,but current knowledge distillation methods mostly target classification tasks,target detection tasks,segmentation tasks etc.These knowledge distillation methods with significant performance can not be well adapted to the person re-identification model,and the key features of person re-identification accuracy are not optimized.We researche and improve the above problems:(1)We are committed to obtaining a better feature representation and feature learning,discovering a richer person features,combining global and local information,and exploring an effective dual-branch network structure(PGAT).We first combine global and local representations with generalized average pooling(Ge M pooling)to extract richer global and local information;Then,the Additive Angular Margin Loss(Arcface)is introduced in the global branch,and it is jointly trained with the Triplet loss to improve the feature learning ability and obtain better robustness;Finally,the experiment is compared and evaluated on three large data sets.The accuracy of the method proposed in this thesis on the Market-1501,Duke MTMC-re ID and CUHK03 datasets reaches 89.54%,79.86%,and 63.76%,respectively,and has certain competitiveness among the existing methods.(2)Considering the actual task requirements for person re-identification,it is necessary to ensure that the model has a shorter reasoning time and higher recognition accuracy.We propose a multi-stage distillation frame structure suitable for person reidentification models.Distill the features extracted by the model,and perform auxiliary distillation at multiple stages of the model backbone network at the same time,improving the recognition accuracy of the PCB network(student network)whose backbone network is Res Net18 to a level equivalent to that of Res Net50(teacher network).The accuracy of the student network Res Net18 on the Market1501,Duke MTMC-re ID,and CUHK03 datasets increased by 4.21%,5.8%,and 5.82%,respectively,compared to before distillation.At the same time,the multi-stage distillation method is applied on the newly proposed dual-branch network(PGAT)to verify the effectiveness of the distillation method proposed in this thesis in the person re-identification model.