Memory-augmented graph convolutional network for vision ship recognition in intelligent maritime surveillance systems

In harsh weather, uneven lighting, and complex nearshore environments, static visible light ship recognition in intelligent maritime surveillance faces numerous challenges. These conditions obscure fine-grained target details, and existing methods often lose shallow information during network propagation. This makes small-scale ships difficult to recognize. To address this issue, we propose a memory-augmented graph convolutional network (MAGCN) for ship recognition. The proposed adaptive feature selection (AFS) module progressively extracts discriminative entity nodes from multi-scale Swin-Transformer features, and the GCN memory (GM) module re-lates these nodes to fuse shallow and deep contextual information recur-sively, thereby preserving critical fine-grained details. Experimental results on CIB-ships, MAR-ships, and Game-of-ships show that MAGCN improves accuracy by 2.65% on CIB-ships, is 0.41% lower than FREGNet on MAR-ships, and exceeds FREGNet by 0.02% on Game-of-ships. Paired tests over five fixed random seeds show reproducible improvements over the internal Swin-Transformer-Base+MLP baseline, but do not establish statistical su-periority over the external competing methods. These results indicate that the advantage of MAGCN is most evident in the highly cluttered CIB-ships setting, while its performance on the other two datasets is competitive rather than uniformly superior.