Global-to-Local Hierarchical Representation Learning for Maritime Vessel Re-Identification

Vessel Re-Identification (Re-ID) plays an important role in the fields of maritime surveillance and traffic flow analysis. However, complex background influence and viewpoint/scale variations jointly make it difficult to accurately identify instances. Focused on this, we propose a Global-to-Local Hierarchical Representation Learning (GLHR) method for maritime vessel Re-ID, in this paper. Here, a unified two-stage framework is designed, where hierarchical feature representation is first learnt, followed by global-to-local aligned re-ranking to further optimize the identification results. First, to construct a robust global representation while preserving discriminative local details, we introduce a Hierarchical Feature Aggregation (HFA) module at the representation learning stage. The HFA coordinates features from different network stages and aggregates them into a unified descriptor for vessels. This design enforces cross-scale geometric consistency between holistic contextual representations and local neighborhood patterns. Then, we develop a Global-to-Local Structure-Aligned Re-ranking (GLSA) strategy by computing local alignment-aware similarity and incorporating it into the global retrieval distance. Specifically, we propose an efficient Top-k Patch Matching (TPM) algorithm to establish sparse correspondences between image pairs with a bidirectional Chamfer distance-based scheme. These sparse matches effectively suppress background interference and alleviate the misalignment of local relationships caused by perspective changes. Experimental results on public vessel Re-ID datasets and the self-constructed dataset demonstrate the effectiveness and superiority of the GLHR in improving vessel Re-Identification accuracies. The related code will be publicly available at https://github.com/kuofku/GLHR.