Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (3): 215-221.DOI: 10.3778/j.issn.1002-8331.2009-0082

• Pattern Recognition and Artificial Intelligence • Previous Articles     Next Articles

Pedestrian Re-identification Combining Random Erasing and Residual Attention Network

SHE Xiangyang, LI Ruixin, YE Ou   

  1. College of Computer Science and Technology, Xi’an University of Science and Technology, Xi’an 710054, China
  • Online:2022-02-01 Published:2022-01-28

融合随机擦除和残差注意力网络的行人重识别

厍向阳,李蕊心,叶鸥   

  1. 西安科技大学 计算机科学与技术学院,西安 710054

Abstract: Traditional pedestrian re-recognition methods rely on artificially constructed visual features, which are easily affected by other external factors and have low recognition accuracy. The deep learning model can extract features autonomously, but as the number of network layers deepens, the gradient disappears. The residual network can alleviate the gradient disappearance problem, but the extracted feature information is not used rationally. Partial occlusion of pedestrian images is another important factor affecting the accuracy of pedestrian re-identification. To solve the above problems, this paper proposes a pedestrian re-recognition algorithm combining random erasing and residual attention network. The algorithm:first, on the basis of the residual network, the attention mechanism module is introduced, and the discriminative ability of the network is improved by strengthening the useful features and the features with little inhibition. Second, introduce random erasing data enhancement method in order to reduce the over-fitting phenomenon, at the same time improve the network generalization ability, and solve the occlusion problem in pedestrian re-identification. Third, using triplet loss to supervise and train the fusion network to achieve better clustering effect of samples in the feature space and improve the accuracy of pedestrian re-recognition. Experiments show that the algorithm can obtain higher recognition accuracy on the Market-1501 dataset and DukeMTMC-reID dataset.

Key words: pedestrian re-identification, random erasing, residual network, attention mechanism, deep learning

摘要: 传统的行人重识别方法依赖人工构造视觉特征,容易受到其他外界因素的影响,识别精度低。深度学习模型能自主地提取特征,但随着网络层数的加深会出现梯度消失情况,残差网络能缓解梯度消失问题,但提取出的特征信息未被合理使用。行人部分图像被遮挡是影响行人重识别准确性的另一个重要因素。针对上述问题提出了融合随机擦除和残差注意力网络的行人重识别算法。该算法:(1)在残差网络的基础上,引入注意力机制模块,通过强化有用的特征和抑制作用不大的特征来提升网络的判别能力。(2)引入随机擦除的数据增强方法,以便降低过拟合现象,同时提高网络泛化能力,解决行人重识别中遮挡问题。(3)使用triplet loss对融合网络进行监督训练,实现样本在特征空间中达到更好的聚类效果,提升行人重识别的准确率。实验表明,该算法在Market-1501数据集和DukeMTMC-reID数据集上能获取较高的识别精度。

关键词: 行人重识别, 随机擦除, 残差网络, 注意力机制, 深度学习