基于注意力机制修正网络的行人属性识别

针对现有行人属性识别方法模型复杂,识别性能较低的问题,提出一种端到端的行人属性识别方法。构建注意力机制修正网络,在主干网络的不同卷积层后添加注意力分支,以提取注意力特征关注属性相关空域;提出一种注意力机制辅助训练方法,将注意力分支与主网络在预测级进行损失融合,通过梯度反向传播修正主网络权重,实现主网络的有效训练;在预测阶段,利用权重修正后的主网络实现属性识别。在RAP数据集上的实验结果表明,提出方法在没有额外辅助信息、不增加主网络体积和计算量的情况下,提升了行人属性识别性能。

Pedestrian attribute recognition based on attention mechanism refined network

Aiming at the problems that the models of pedestrian attribute recognition are complex and the performance still needs to be improved,an end-to-end method was proposed.An attention mechanism refined network was constructed.Several attention branches were added behind different convolutional layers of the main network to extract attention features which focused on the attention-related spatial area.An attention-mechanism-assisted training method was proposed,the losses of attention branches and main network were aggregated at prediction level.The weights of the main network were refined by attention branches via error back propagation,and the main network was trained more effectively.In prediction stage,the refined main network was used for attribute recognition.Experimental results on RAP dataset show that,the proposed method can effectively improve the pedestrian attribute recognition performance without increasing the volume and calculation of the main network and additional information.