Multiscale attention dynamic aware network for fine-grained visual categorization

Fine-grained visual categorization (FGVC) is a challenging task, facing the issues such as inter-class similarities, large intra-class variances, scale variation, and angle variation. To address these issues, the authors propose a novel multiscale attention dynamic aware network (MADA-Net). The core of network consists of three parallel sub-networks, which learn features from different scales. Each sub-network is composed of three serial sub-modules: (1) A self-attention module (SAM) locates objects according to relative importance scattered throughout feature map. (2) A multiscale feature extractor (MFE) learns the non-linear features of objects. (3) A dynamic aware module (DAM) enhances the learning capability of spatial deformation of the network to generate high-quality feature map. In addition, the authors propose a multiscale adjusted loss (MA-Loss) to improve the performance of network. Experiments on three prevailing benchmark datasets demonstrate that our method can achieve state-of-the-art performance.