Due to the limited storage and computing resources of underwater detection instruments, using high-parameter models will occupy more storage space and computing power. To solve this problem, a target detection model based on YOLOv8n, called YOLOv8n-MAL, was proposed, which achieves lightweight overall while maintaining detection accuracy. Firstly, a lightweight multi-scale convolutional block attention module (MSCBAM) is proposed, which can extract features of different scales from input feature graphs through convolution kernel and pooling operations of different sizes. It can improve the robustness of the model in complex scenarios and ensure that the model can maintain high detection accuracy in the face of different types of inputs. Then, a new neck model multi-scale feature fusion network (MSFFN) is designed. Compared with the original neck model, the multi-scale fusion capability of the model is enhanced, and the interaction of features at different levels is strengthened, so that the high-level and low-level features can be more fully combined. This cross-level fusion can make use of the feature representation ability of the network more effectively, avoid the loss or weakening of information in the process of transmission, and improve the detection effect of the model. Secondly, a lightweight multi-scale convolution module lightweight multi-scale convolution module (LMSCM) is proposed, and the module and some convolution modules are integrated into the C2F module to form PC2F-LMS. By introducing a more efficient convolution structure. Finally, the original network loss function is optimized using WIoU. The experimental results show that the average accuracy mAP@0.5 of the improved algorithm on the URPC dataset is increased by 1.4%, and the number of parameters is reduced by 38.6% compared with the YOLOv8n algorithm, which provides an effective reference value for underwater target detection.