Underwater target detection technology has received extensive attention as a crucial auxiliary cultivation technique in the field of intelligent aquaculture. However, traditional target detection algorithms face challenges in maintaining detection accuracy and robustness in underwater applications due to the complexity of the underwater environment and the degradation of underwater optical images. To address these issues, an improved YOLOv8 (You Only Look Once) optimization method specifically designed for underwater target detection is proposed. This method replaces the original C2f module with dynamic snake convolution (DSConv) to adapt to situations where targets and backgrounds overlap in underwater images. Furthermore, continuity constraints are introduced into the design of convolution kernels to enhance perception capabilities. The previous convolution position is utilized as a reference, allowing freedom to select swing directions to determine each convolution position. Experimental results demonstrate that the enhanced network model exhibits favorable performance in underwater target detection, achieving a precision of 83.0% and a recall rate of 74.8%. It outperforms Faster R-CNN, YOLOv3, YOLOv5, YOLOv7, and YOLOv8 networks.