Shift Transformer In Transformer Network for Chinese Ship License Recognition

In intelligent waterway transportation supervision systems, ship license plate recognition plays a pivotal role in maritime safety management and regulatory enforcement. Unlike standardized vehicle license plates, Chinese ship license plate recognition faces multiple technical challenges: complex water surface reflections, variable capture angles, nonuniform font specifications, and particularly the multi-level structural characteristics unique to Chinese characters (comprising radicals and whole characters). To address these issues, this study proposes a Chinese ship license plate recognition network based on Pyramid Vision Transformer architecture. We innovatively design the Shift Transformer in Transformer (STnT) Network, which integrates local and global self-attention mechanisms to achieve collaborative modeling of holistic character structures and radical-level features. Leveraging the structural properties of Chinese characters, we construct a dual-stream multi-task learning framework: the main feature stream employs global context analysis to decode the holistic character semantics, while the auxiliary feature stream utilizes a combination of local and global features to capture radical-level information. Experimental results on the CSLD dataset demonstrate that our method achieves 92.78% character recognition accuracy at 525 FPS, significantly outperforming existing approaches. With a compact parameter size of 9.43 M, this solution provides an efficient and reliable technical foundation for intelligent maritime supervision systems.