For person re-identification,attention mechanism has the advantages of enhancing salient features and suppressing irrelevant features,while previous attention-based methods mostly focus on channel and spatial independently and ignore the cor-respondence relationship between channel and space.To solve this problem,a person re-identification method of dimension fusion attention was proposed.During the feature extraction of the network,the sliding window method was used to fuse the channel dimension and the space dimension,and the super-large one-dimensional convolution kernel was used to continuously learn the relationship between the channel and the space.ECA attention was introduced in the final stage of the network.ECA attention has the effects of local cross-channel interaction.When used in combination with dimensional fusion attention,it significantly improved the recognition rate.Experimental results show that the method is superior to most current methods in the case of limi-ted computational cost.
关键词
行人重识别/维度交互/维度融合/深度学习/卷积神经网络/注意力机制/轻量级
Key words
person re-identification/dimensional interaction/dimensional fusion/deep learning/convolution neural network/attention mechanism/lightweight
维普
万方数据