Research On Person Re-identification Based On Multi-task Joint Supervised Learning

Person re-identification technology,as one of the key research contents in video analysis,has attracted a large number of scholars' attention.Person re-identification mainly refers to using person visual features to determine whether the persons appearing under different cameras are the same person.At present,the research of person re-identification focuses on the extraction features of person.The method of feature extraction algorithms of person re-identification based on manually extracting features of person focuses on the processing of low-level features such as color and texture.The research on feature extraction algorithms of person re-identification based on deep learning is mainly aimed at person verification or identification task,by constructing complex neural network structures to match person's features.Extracting more discriminative and more robustness person features by designing a simple and effective network structure is the key work of this paper on the research of person re-identification.This paper proposes a network structure based on multi-task joint supervised learning,which can improve the accuracy of person re-identification by extracting more distinguishable deep convolution features of person.The multi-task joint supervised learning network structure mainly involves two aspects: the joint supervision of person identification learning and the assisted feature extraction of person multi-attributes learning.This paper proposes to use two loss functions to jointly supervise the person re-identification learning task.It can not only focus on differences between different persons but also learns the correlation information between same persons.We propose to use joint supervision learning method of above two types of information to guide the updating of network model parameters.And this network structure can effectively improve the feature representation performance of person.We also consider that single identification of person has limitation in guiding the network to learn the specific features of person.Therefore,this paper proposes a method by adding person attributes learning tasks based on the joint supervised learning task.This improvement method transforms the original learning objective from simple person identification learning to detailed person descriptive semantic features learning.The semantic features can effectively reduce the impact of the change of person under the cross-camera by obtaining more accurate description of person features.For the realization of the proposed network structure,this paper adopts the multi-task learning method.We use the network sharing structure to realize the correlation learning between multi-tasks,and it can effectively reduce the computational cost of the model.The proposed network structure can obtain more effective features of person images to improve the accuracy of person re-identification.In order to verify the validity of multi-task joint supervised learning network structure proposes in this paper,we conduct experiments on Market1501 database and Duke MTMC-reid database of person re-identification.Comparing with the network structure of single Softmax supervised function for person identification guidance,the method proposes in this paper has improved the accuracy in top-1 and m AP is 8.37% and 13.23% in Market1501 database,respectively.And our method has promoted the accuracy in top-1 and m AP is 10.28% and 9.47% in Duke MTMC-reid database,respectively.The experimental results show that the network structure extraction of deep convolution features based on multi-task joint supervised learning has proposed in this paper can effectively improve the accuracy of person re-identification.