Person Re-identification Base On Contrastive Learning With Attention Mechanism

With the growth of social public security needs,high-definition monitoring equipment has been found in every corner of the city.People’s deeds are saved in the form of image sequence.The traditional mode based on human eye resolution is not effective or efficiency.The task of person re-identification(re-id)is to solve the problem of pedestrian matching under the cross camera equipment.It is usually considered as a sub task of target retrieval,which mainly uses more comprehensive information such as pedestrian gait,body characteristics to identify people.The algorithm of Re-id based on deep learning has become a hot research direction.Contrastive learning is a method of learning by using contrastive loss,which can be seen as a process of looking up a dictionary.The effect of learning is often proportional to the scale of negative sample pairs.In order to increase the number of negative sample pairs in a mini＿batch learning,a contrastive learning siamese network with slowly updated weight parameters is constructed,which can maintain sufficient scale and stability of negative samples,and noise-contrastive estimation loss(NCE)is introduced for training.In order to improve the learning ability of the backbone network to the high-level semantic features,the "slow features" learning from siamese network branch will be filtered and treated as negative samples,so that the backbone network can get rid of the shallow semantic features and explore in the higher-level semantic feature space.In order to reduce the introduction of noise in negative sample filtering,through a cross attention module based on space and channel,we can learn the common feature information of two network branches,so that the "slow feature" filtering is more targeted.In the inference stage,the difference between the two branch networks is fully utilized,and the output of the backbone network and the attention module is taken as the final feature representation.Experiments show that the person re-identification network with contrastive learning has achieved a significant improvement compared with the baseline network.After the integration of attention module,the network representation ability has been further enhanced.Under the premise of using only global features,person re-identification based on attention mechanism with contrastive learning has a certain competitiveness compared with the current mainstream algorithm.