Region-Based Global-Local Contrastive Representation Learning for Fine-Grained Ship Classification in Remote Sensing Images

Fine-grained ship classification (FGSC) is a critical task in both military and civilian domains, aimed at accurately identifying and categorizing sub-classes of ships. However, achieving highly discriminative fine-grained feature representations for FGSC tasks with remote sensing images (RSIs) remains challenging. Existing methods often fail to effectively capture subtle local differences in complex remote sensing scenes, leading to suboptimal performance. To address this problem, we propose a region-based feature learning framework to learn discriminative feature representations from a regional perspective. Specifically, we design a discriminative part discovery network (DPDN) that aims to learn the key sub-regions for each object. At the same time, we promote the diversity within these discriminative regions to obtain informative feature representations, facilitating a comprehensive understanding of objects. To further enhance the discriminativeness of the learned representations, we introduce a global-local contrastive learning (GLCL) strategy. This strategy employs contrastive tasks at both global and local granularities, thereby effectively capturing contrastive cues at multiple levels. We evaluate our method on two widely-used FGSC datasets (i.e., FGSC-23 and FGSCR-42 datasets). The experimental results demonstrate that the proposed method outperforms previous state-of-the-art methods.