Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (8): 191-197.DOI: 10.3778/j.issn.1002-8331.2012-0143

• Pattern Recognition and Artificial Intelligence • Previous Articles     Next Articles

Re-Identification Method of Siberian Tiger Based on Adaptive Regularization

YU Huiling, QIAN Chengshuai   

  1. College of Information and Computer Engineering, Northeast Forestry University, Harbin 150040, China
  • Online:2022-04-15 Published:2022-04-15

基于自适应正则化的东北虎重识别方法

于慧伶,钱成帅   

  1. 东北林业大学 信息与计算机工程学院,哈尔滨 150040

Abstract: As the number of Siberian tigers continues to decrease, it becomes very meaningful to identify a single tiger for protection and tracking. Therefore, a Siberian tiger re-identification network model based on local partial blocks and an adaptive L2 regularization method(part- based convolutional baseline-adaptiveL2, PCB-AL2) is adopted to solve the problem of difficulty in re-identification of Siberian tiger in natural environment. The adaptive L2 regularization factor is adaptively updated through backpropagation, which is achieved by using the regularization factor as a trainable variable. Aiming at the characteristics of tigers relying on body stripes to distinguish, a two-branch network structure is adopted:local branch and global branch. The network relies on local features to guide global feature learning. Experimental results show that comparing with PPbM-a, PPbM-b and PPGNet on the ATRW data set, it is concluded that mAP reaches 92.1% in a single-camera environment and 75.1% in a cross-camera environment.

Key words: re-identification, residual network, adaptive L2 regularization, characteristics of fusion

摘要: 随着东北虎数量不断减少,识别单只老虎进而做出保护和追踪变得很有意义,故采用了一种基于局部分块和自适应L2正则化方法的东北虎重识别网络模型(part-based convolutional baseline-adaptiveL2,PCB-AL2)以解决在自然环境下东北虎重识别困难等问题。自适应L2正则化因子通过反向传播进行自适应更新,这是通过将正则化因子作为可训练的变量来实现的。针对老虎依靠身体条纹分辨的特点,采用一种双分支网络结构:局部分支和全局分支,网络依靠局部特征指导全局特征学习。实验结果表明,在ATRW数据集上与PPbM-a、PPbM-b以及PPGNet对比得出结论,在单摄像头环境下mAP达到了92.1%,跨摄像头环境下mAP达到75.1%。

关键词: 重识别, 残差网络, 自适应L2正则化, 特征融合