| Person re-identification aims to find specific person by matching person images and jointly deploying non-overlapping surveillance cameras at different locations.It has very important applications in video surveillance and other fields.Using attention mechanism to enhance deep neural networks provides an effective parallelizable feature extraction method,and it has been well applied in many computer vision related tasks.This paper has conducted in-depth analysis and research on person re-identification based on attention mechanism.The main work and results are as follows:(1)In this paper,a Cross-level Reinforced Attention CNN(CLRA-CNN)is proposed to solve the problem that most of the existing attention based methods modeling attention relationship only for the current level features,while meaningless information still occupies a certain weight that causes interference to the model training.Firstly,a Cross-level Feature Fusion module(CLFF)is designed to adaptively fuse features of different levels to guide subsequent attention module generation.Secondly,a Reinforced Attention module(RA)is designed by combining soft attention with hard attention to adjust the weight of features in space and channel,respectively.And then,it can obtain more discriminative weighted reinforcement features.Further,by combining the CLFF module with the RA module,the Cross-level Reinforced Attention(CLRA)is obtained.Through the CLRA module,the network can aggregate and propagate more discriminative semantic information to optimize the person re-identification task.(2)In this paper,a Multi-scale Compressed Reinforced Attention CNN(MSCRA-CNN)is proposed to solve the problem that most attention methods usually generate attention maps only by fixed-scale features.And it can achieve weighted enhancement for multi-scale features.Firstly,a Multi-scale Feature Fusion module(MSFF)is designed by combining the original scale feature with the larger feature of the receptive field size.This model can adjust the receptive field size adaptively according to the input.Furthermore,in order to improve the realtime performance of RA,this paper proposes a new Compressed Reinforced Attention(CRA)to optimize the attention from space and channel,respectively.By combining the maximum pooling and average pooling,the compressed features can be realized.And the compressed features are used to model the interdependence between spatial positions or channels.Finally,by combining the MSFF module with the CRA,the Multi-scale Compressed Reinforced Attention(MSCRA)is proposed to extract more discriminative features.(3)In this paper,a Multi-scale Attention Multi-branch Network(MSAMB-CNN)is proposed.Most existing deep learning networks pay more attention to the most significant information in person images,but pay less attention to the sub-significant details.The proposed method can achieve the coordination of the image significant information and sub-significant information.Firstly,the proposed method combines the simplified multi-scale feature fusion method with the dual attention mechanism to design a Multi-scale Attention module(MSA).through the interactive fusion of different-scale features,the network can adaptively adjust the receptive field size according to the input information.Secondly,a weighted feature dropout method is obtained by combining the MSA module with the feature discarding method to realize the weighted reinforcement for the local features.Further,a fuzzy image partition strategy method is proposed.It can combine with weighted feature dropout to obtain more detailed and abundant local features.Finally,a multi-branch network is established to realize the coordination and unification of global features and multiple local features.The proposed method can realize the weighted strengthening of global salient information and sub-significant local detail information respectively and obtain more discriminative features for final identification.In summary,this paper proposes three attention-based person re-identification networks: CLRA-CNN,MSCRA-CNN and MSAMB-CNN,and demonstrates the excellent performance of the proposed algorithm through a large number of experiments on several public datasets. |
