To address the issues of limited computing power of underwater equipment and low clarity of underwater images, an improved lightweight YOLOv8 algorithm is proposed. First, the Cross-scale Convolutional Feature-fusion Module (CCFM) is introduced to improve the model’s performance in dealing with multi-scale underwater targets. The CCFM enhances the detection accuracy of small targets while reducing the number of parameters and computation. Then, the detection performance and efficiency is improved by introducing a dynamic head to unify the task-awareness, scale-awareness, and spatial-awareness. The dynamic head effectively enhances the clarity of images. Subsequently, the Mixed Local Channel Attention (MLCA) is introduced. MLCA enhances the network’s ability to extract key features while ensuring the computational and detection efficiency of the model. The experimental results show that compared with the original model (yolov8n) on the publicly available underwater target detection dataset RUOD without using pre-trained weights. The following is an analysis of the data. The map50 reaches 85% and improves by 0.8%, the map50-95 improves by 0.9%, the amount of parameters is reduced by 21.4%, the amount of computation is reduced to 7.4 GFLOPs, and the size of the model is reduced to 5.1M. In this paper, the original yolov8 is lightened as well as the accuracy is improved, and the improved algorithm is well suited for target detection in underwater robots.