Underwater images are affected by factors such as complex marine environments and many interfering objects, leading to the fact that existing underwater target detection algorithms often face problems of low detection accuracy, false detections and missed detections. In addition, the underwater target detection model is complex, while the exploration equipment has limited memory, high deployment cost and poor real-time detection performance. To solve these problems, a lightweight underwater target detection algorithm DAF-YOLO based on improved YOLOv8s is proposed.The algorithm improves the backbone network of YOLOv8s, and designs an improved C2f module D2F, which enhances the feature characterisation capability, better captures complex details and handles deformations, and solves the problem of misdetections and miss-detections in detection. Meanwhile, to address the semantic gap and information conflict when fusing non-neighbouring layers in the neck feature fusion network, an asymptotic feature pyramid network, AFPN, is used to support the direct interaction of non-neighbouring layers, to achieve the full fusion of multi-scale features, and to enhance the feature fusion capability of the model so as to improve the target detection accuracy. In order to reduce the computational burden of the neck network, a novel convolutional approach of PConv in FasterNet is also introduced, which reduces the number of parameters and the computational volume, achieves the lightweight of the neck network, and optimally improves the detection speed of the model. Experimental tests show that the model achieves a 3.9% improvement in accuracy on the URPC2020 dataset, while achieving high-speed detection at 115.6 frames per second. In addition, the parameters and computation of the model are significantly reduced by 10.52% and 10.91%, respectively, which further achieves the balance between accuracy and speed of the algorithm.