Multi-scale discriminative regions attention network for fine-grained vehicle classification

作     者：Rong, Wen-Zhong Han, Jin Cai, Ying-Hao Liu, Gen 

作者机构：College of Computer Science and Engineering Shandong University of Science and Technology Qingdao266590 China The State Key Laboratory of Management and Control for Complex Systems Institute of Automation Chinese Academy of Sciences Beijing100190 China 

出 版 物：《Journal of Network Intelligence》 (J. Network Intell.)

年 卷 期：2021年第6卷第2期

页      面：164-177页

核心收录：

基　　金：Acknowledgment. This work is supported by National Natural Science Foundation of China (U1913201)  The Key Project of Shandong Provincial Natural Science Foundation (ZR2020KE023)  Collaborative Education Project of Ministry of Education (201901055015)  Postgraduate Education Quality Improvement Project of Educational Commission of Shandong Province of China (ADYAL17034)  and Excellent Teaching Team Support Project of Shandong University of Science and Technology (JXTD20170503) 

主　　题：Classification (of information) 

摘      要：Fine-grained vehicle classification is a challenging task in computer vision due to the low intra-class variance. Some methods have been developed to improve the accuracy of fine-grained vehicle classification by improving the ability of discriminative features extraction, but there is still room for further improvement in the localization accuracy of vehicle discriminative regions. Based on deep convolutional neural networks, we focus on finding a more efficient structure that pays more attention to the discriminative image regions to enhance the ability of fine-grained vehicle classification. We propose a novel Multi-Scale Discriminative Regions Attention Network (MS-DRAN), which extracts feature maps through ResNet-50 backbone network and generates multi-scale feature maps by a Feature Pyramid Network (FPN). Then, MS-DRAN generates discriminative regions attention maps on the multi-scale feature maps. The attention maps from shallow layer perform pixel-level multiplication with the feature maps from deeper layer. By this way, the network gradually extracts a more discriminating feature maps for classification. We also design a multi-task loss function to associate the classification results for each scale and optimize the network parameter in training. Lastly, we validate the MS-DRAN on Stanford Cars-196 dataset and CompCars dataset and achieves 94.3% and 98.1% in accuracy, respectively. ©2021 ISSN 2414-8105 (Online).