Coarse-to-Fine Granularity in MultiScale FeatureFusion Network for SAR Ship Classification

Wei Lin, Hao Zheng, Zhigang Hu, Meiguang Zheng, Liu Yang

Abstract: The classification of Synthetic Aperture Radar (SAR) ships is a challenging task due to the small inter-class differences and large intra-class variance. Previous methods have used multiscale feature fusion to solve this problem, but most of them are too rough to extract fine-grained features. In this paper, we propose a new coarse-to-fine granularity multiscale feature fusion network (C2F-MFF) to address this issue. C2F-MFF consists of two stages, i.e., coarse-grained multiscale feature extraction and adaptive fine-grained multiscale feature refining. The first stage is used to capture and augment the discriminative scale-rich fusion features, while the second stage is able to adaptively assign varying importance to individual scale features. Notably, the first stage contains two novel blocks, feature focus (FF) block and feature enhance (FE) block are interactively introduced to capture significant and abundant information. Extensive ablation studies can confirm the effectiveness of each contribution. Results on the three-category and six-category OpenSARShip datasets demonstrate that our network surpasses the modern CNN-based methods and other feature fusion methods.