Vessel re-identification in maritime surveillance presents substantial challenges, primarily stemming from significant appearance variations caused by diverse viewpoints and the limited availability of annotated data. Effectively aggregating features that are both view-adaptive and discriminative is essential for achieving robust performance. To address these issues, we propose a novel framework named View-invariant Adaptive Multi-granularity Network (VAMN), which is designed to extract identity-invariant features from multi-view and environmentally varying vessel images. The proposed VAMN architecture consists of three key components. First, the Adaptive Mixed Pooling (AMP) module adaptively balances average pooling and max pooling operations via a lightwe-ight network. Second, the View-invariant Feature Learning (VFL) module enhances feature robustness through multi-frequency phase modulation, explicitly modeling viewpoint variations. Third, the Multi-granula-rity Pseudo-label Generation (MPG) strategy exploits adaptive representations to generate reliable pseudo-labels. The tight integration of these components forms a positive feedback mechanism that progressively improves feature discrimination and alignment. Extensive experiments conducted on the VesselReID dataset demonstrate that VAMN achieves state-of-the-art performance, underscoring its overall effectiveness and the complementary advantages of each proposed module.