Multi-Scale Adaptive-Indexed Scanning Vision Mamba Guided Lighten YOLO-v5 for Coal Gangue Separation

Due to the different sizes, complex backgrounds, variable lighting, and severe occlusion, the detection of small coal gangue targets is a challenging task. Traditional YOLO-v5 based methods exhibit low accuracy in such scenarios. Therefore, we propose an improved YOLO-v5-based model that incorporates the RepNCSPELAN4 architecture from YOLO-v9 and the enhanced Vision-Mamba architecture with multi-scale adaptive position indexing. This approach is designed to better localize coal gangue targets of varying sizes. Experimental results indicate that our model reduces the parameter count to 0.64M and decreases the number of floating-point operations while achieving a detection accuracy of 98.8%. This demonstrates significant practical value for real-world applications, particularly in the context of detecting small targets under complex conditions.