融合多分支与多粒度特征的东北虎重识别  

Re-Identification of Amur Tigers by Combining Multi-branch and Multi-granularity Features

作　　者：李晓楠 朱朦 任洪娥[1,3] 陶锐 LI Xiaonan;ZHU Meng;REN Honge;TAO Rui(College of Information and Computer Engineering,Northeast Forestry University,Harbin 150040,China;Harbin University,Harbin 150086,China;Heilongjiang Forestry Intelligent Equipment Engineering Research Center,Harbin 150040,China;Hulunbuir University,Hulunbuir 021008,China)

机构地区：[1]东北林业大学信息与计算机工程学院,哈尔滨150040 [2]哈尔滨学院,哈尔滨150086 [3]黑龙江林业智能装备工程研究中心,哈尔滨150040 [4]呼伦贝尔学院,内蒙古呼伦贝尔021008

出　　处：《哈尔滨理工大学学报》2023年第6期95-102,共8页Journal of Harbin University of Science and Technology

基　　金：黑龙江省自然科学基金项目(LH2020F040);中央高校基本科研业务费专项资金资助项目(No.2572017PZ10).

摘　　要：为解决东北虎重识别研究中存在的细节特征提取不充分等问题,提出了一种融合多分支与多粒度特征的东北虎重识别模型CMM-Net。其中,全局分支负责提取宏观上的粗粒度特征;注意力分支通过插入坐标注意力模块加深了网络对重要特征的关注度;局部分支通过将特征图切分成不同条带块,从而提取东北虎更细粒度的局部特征。通过多个分支结构和多个细粒度特征结合来对模型进行优化学习,加强全局特征与局部特征的关联性。同时提出用Circle Loss与Softmax的联合损失来提高网络精度。实验结果表明,在ATRW数据集上所提模型在单摄像头环境下mAP为93.6%,跨摄像头环境下mAP为77.4%,均优于多数文献所提方法,证明了本文模型的有效性。In order to solve the problem of insufficient detailed feature extraction in the re-identification of the Amur tiger,a re-identification model of the Amur tiger,CMM-NET,was proposed,which combined multi-branch and multi-granularity features.The global branch is responsible for extracting macroscopic coarse-grained features.The attention branch deepens the network′s attention to important features by inserting coordinate attention module.Local branches can extract finer grained local features of amur tigers by cutting the feature map into different blocks.Finally,the model is optimized by combining multiple branch structures and multiple fine-grained features to strengthen the correlation between global features and local features.Meanwhile,the combined Loss of Circle Loss and Softmax is proposed to improve the network accuracy.Experimental results show that the mAP of the model proposed on THE ATRW data set is 93.6%in a single camera environment and 77.4%in a cross-camera environment,both of which are better than the methods proposed in most literatures,proving the effectiveness of the proposed model.

关 键 词：东北虎重识别 残差网络 多分支特征 坐标注意力机制 circle loss 

分 类 号：TP751.1[自动化与计算机技术—检测技术与自动化装置]