结合指导学习和特征擦除的东北虎重识别研究  

Research on Re-recognition of Amur Tiger Based on Guided Learning and Feature Erasure

作　　者：翟冰 戴天虹[1] ZHAI Bing;DAI Tianhong(College of Mechanical and Electrical Engineering,Northeast Forestry University,Harbin,150040,China)

机构地区：[1]东北林业大学机电工程学院,哈尔滨150040

出　　处：《野生动物学报》2023年第2期297-304,共8页CHINESE JOURNAL OF WILDLIFE

基　　金：黑龙江省自然科学基金项目(C201414);中央高校基本科研业务费专项资金项目(2572019CP17);哈尔滨市科技创新人才项目(2014RFXXJ086)。

摘　　要：东北虎(Panthera tigris altaica)重识别对于东北虎的行为研究以及保护具有重要意义,针对在自然环境中东北虎难以被重新识别的问题,提出了一种结合指导学习和特征擦除的东北虎重识别方法。设计了一个包括全局流和局部流的双流网络,局部流以分割后的前景图作为输入,使网络更关注前景部分,全局流以原图作为输入,通过局部流指导全局流学习细粒度特征,同时将模型最后全连接层(FC)的各类别权重参数之间的差异作为指导进行特征擦除,以擦除后的特征向量训练网络,从而增强特征向量中更多元素的辨别能力,最后在测试阶段只使用全局流。在野外东北虎重识别(ATRW)数据集上对该方法进行评估,在单摄像机条件下,m AP达到92.1%;在交叉摄像机条件下,m AP达到72.9%,试验结果表明方法有效。Re-recognition of Amur tiger(Panthera tigris altaica)is of great significance to the behavioral studies and Amur tiger conservation.Aiming at the problem that it is difficult to re-recognize Amur tiger in the natural environment,a rerecognition method combining guided learning and feature erasure is proposed.A dual flow network including global flow and local flow is designed.The local flow takes the segmented foreground image as the input,so that the network pays more attention to the foreground part.The global flow takes the original image as the input,and the local flow guides the global flow to learn fine-grained features.At the same time,the difference among various weight parameters of the full con⁃nection layer(FC)at the end of the model is used as the guidance to erase the feature,and the erased feature vector is used to train the network.Thus,the ability to identify great amount of elements in the feature vector is enhanced.Finally,only the global flow is used in the test phase.The method was evaluated on the ATRW dataset in the wild.Under the condi⁃tion of single camera,the mAP reached 92.1%;and under the condition of cross camera,the mAP reached 72.9%.The experimental results show that the method is effective.

关 键 词：东北虎重识别 双流网络 细粒度特征 特征擦除 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] Q958.1[自动化与计算机技术—计算机科学与技术]