Evolutionary Tuner and Selective Kernel Attention for Improving YOLOv11 in Underwater Fish Detection and Recognition

This study aims to enhance the accuracy of underwater fish detection by proposing a dual-enhanced YOLOv11 model. The approach leverages two key improvements: First, a selective kernel attention (SKA) mechanism is incorporated into the YOLOv11 architecture to enable dynamic selection of multi-scale convolution kernels, improving adaptability to various target sizes. Second, an evolutionary tuner (ET) is employed for hyperparameter optimization to refine model performance further. The proposed model achieves significant gains over the baseline, with improvements of 2.06% in mean average precision (mAP)@0.5 and 6.30% in mAP@0.5:0.95, attaining final scores of 98.629% and 86.933%, respectively. The dual-enhanced model demonstrates superior accuracy and robustness in complex underwater environments, ultimately achieving a precision of 99.069% and a recall of 95.968%.