基于动态卷积与注意力的多特征融合行人重识别

Person re-identification based on multi-feature fusion of dynamic convolution and attention

为能够准确利用图像中有效特征,提取判别性较高的信息区分特征相近的行人,提出一种基于动态卷积和注意力机制的多分支网络。将动态卷积核作用于ResNet50网络中,使动态卷积中的注意力机制与网络中的通道和空间注意力共同作用,通过不同分支得到相应局部特征,融合得到高判别性特征进行分类匹配。在CUHK03、DuckMTMC-reID、Market-1501数据集上进行验证实验,其结果表明了所提模型的优越性。

To accurately use the effective features of the image,extract more discriminative information to distinguish pedestrians with similar features,a multi-branch network based on dynamic convolution and attention mechanism was proposed.The dynamic convolution kernel was applied to the ResNet50 network,so that the attention mechanism in dynamic convolution and the channel and spatial attention in the network worked together,corresponding local features were obtained through different branches,and high discriminative features were obtained by fusion for classification and matching.Verification experiments were carried out on the CUHK03,DuckMTMC-reID,and Market-1501 data sets.Results show the superiority of the proposed model.