YOLO-RDM: A Lightweight and High-Precision Method for Underwater Small Target Detection

Underwater target detection faces great challenges due to low visibility, low contrast and light absorption and scattering. The existing YOLO series has obvious defects when used underwater: the weight of convolutional kernel parameters is fixed, and it is difficult to adapt to complex underwater background and multi-scale targets; The traditional multi-scale feature extraction mechanism can not effectively deal with the problem of fuzzy and dense distribution of small underwater targets. Lack of robust attention mechanism for underwater low-quality images, susceptible to background interference and sensitive to occluded targets. Therefore, the improved model YOLO-RDM based on YOLOv8 is proposed in this study. The RFAConv module, which can dynamically adjust convolution ker-nel parameters, is introduced to strengthen the feature extraction capability of complex underwater scenes. The Dual_ C2f structure with <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$3\times 3$</tex> convolution kernel and IxI convolution kernel is used to optimize the feature interaction, which reduces the computational cost while maintaining multi-scale information fusion. A new multi-scale spatial attention enhancement mechanism, MultiSEAM, is designed to improve the robustness of detection of small targets and occluded scenes through depth-separable convolution and multi-scale feature weighting. In the URPC2020 dataset, [email protected] and [email protected]:0.95 of YOLO-RDM reach 84.0% and 49.8%, respectively, the reasoning speed is 49.9, and the parameter number is only 3.0M. Significantly better than the mainstream YOLO series and RT-DETR and other latest models. The ablation experiments further confirm the effectiveness of the improved modules. The YOLO-RDM achieves a good balance between precision and lightweight, providing an efficient and reliable solution for small underwater target detection.