AMLNet: Attention Multibranch Loss CNN Models for Fine-Grained Vehicle Recognition

Hongchun Lu, Min Han, Chaoqing Wang, Junlong Cheng

Abstract: For fine-grained vehicle recognition tasks, small interclass variation and large intraclass variation are key factors that can degrade recognition performance. Therefore, mining fine-grained discriminative features is an important way to improve performance. Although previous studies have achieved some successes, the standard model still has some drawbacks. First, global features are ignored when extracting discriminative features. Second, the potential interactions between different regions are not sufficiently considered. To address this challenge, we propose an attention multibranch loss convolutional neural network (AMLNet) in a weakly supervised manner to learn features from different regions and explore the interactions between them. Specifically, a pretrained model is used to initially extract rich vehicle feature information, and attention is enhanced using a multifeature fusion global cross-channel loss network (MGCL) and cross-channel pooling enhanced attention feature network (CCPA) to mine discriminative features while enhancing global feature representation. We conduct comparative experiments on two public datasets, and experimental results show that the proposed method achieves better recognition performance than other advanced algorithms. Thus, the proposed method is more effective when applied to fine-grained vehicle recognition tasks.