Adaptive Mid-Level Feature Attention Learning for Fine-Grained Ship Classification in Optical Remote Sensing Images

Xi Yang,Zilong Zeng,Dong Yang

DOI: https://doi.org/10.1109/tgrs.2024.3351874

IF: 8.2

2024-02-02

IEEE Transactions on Geoscience and Remote Sensing

Abstract:Ship classification in optical remote sensing images is a critical task for various maritime applications, including anti-smuggling, maritime traffic control, and maritime rescue. However, fine-grained ship classification (FGSC) is challenging due to the complex background, intraclass similarity, and interclass difference. In this article, we propose a novel mid-level feature attention learning method for FGSC. Our method incorporates mid-level feature casual attention (MFCA) and mid-level channel attention (MCA) to identify discriminative regions and local features corresponding to subtle visual features. The MFCA constrains the learning process of mid-level features through comparison with attention maps and counterfactual attention maps, while the MCA uses a discriminative component to extract discriminative features from channel information and a diversity component to focus feature channels on more obvious feature regions. Besides, an adaptive weight is added to dynamically adjust the influence of MFCA and MCA in the model. Our method can be trained end-to-end and requires no annotations other than category information. Extensive experiments on two large-scale FGSC datasets, FGSC-23 and FGSCR-42, demonstrate that the proposed method achieves state-of-the-art performance, outperforming existing methods by a significant margin.

imaging science & photographic technology,remote sensing,engineering, electrical & electronic,geochemistry & geophysics