FSLNet: Leveraging Skeleton Cue Guided FBIC for Bionic Intelligent Surveillance Robot via Transformer

Fine-grained bird image classification (FBIC) plays a critical role in robotic vision tracking and monitoring of endangered bird species. Traditional surveillance robots often face major challenges when operating in complex environments, especially under extreme conditions involving occlusions and complex backgrounds. To overcome these limitations, we developed a novel bionic intelligent surveillance robot that incorporates FSLNet - a fine-grained skeletal learning network that integrates skeletal tokens into a Vision Transformer (ViT) framework. Unlike traditional CNN-based models, FSLNet leverages skeletal tokens to capture key invariant features, improving the ability to distinguish between visually similar bird species under challenging circumstances. The architecture of our model includes visual token construction, skeleton token construction, and a transformer-based backbone for thorough feature learning. Extensive experiments on CUB-200-2011 and NABirds datasets demonstrate that our method achieves superior performance compared to state-of-the-art approaches.