This paper introduces a lightweight YOLOv8 model, titled UW-YOLOv8, designed for real-time underwater object detection. The method optimizes YOLOv8s by replacing its original backbone with a lightweight FasterNet module to reduce computational load and improve performance. The bi-directional feature pyramid network (BiFPN) is modified to a faster version by reducing unnecessary feature layers and changing the fusion method. Additionally, a lightweight-C2f structure is proposed, which replaces the standard convolution and bottleneck module with a GSCov and partial convolution, respectively, to achieve a lighter and faster block. Experiments on three underwater datasets (RUOD, UTDAC2020, and URPC2022) show that the proposed method achieves mAP50 of 86.8%, 84.3%, and 84.7% respectively, with a speed of 156 FPS on NVIDIA A30 GPUs, meeting real-time detection requirements. Compared to YOLOv8s, the model volume is reduced by an average of 24%, and mAP accuracy is enhanced on all three datasets. The paper also discusses related works in underwater object detection and lightweight network designs, highlighting the balance between model size, speed, and accuracy.This paper introduces a lightweight YOLOv8 model, titled UW-YOLOv8, designed for real-time underwater object detection. The method optimizes YOLOv8s by replacing its original backbone with a lightweight FasterNet module to reduce computational load and improve performance. The bi-directional feature pyramid network (BiFPN) is modified to a faster version by reducing unnecessary feature layers and changing the fusion method. Additionally, a lightweight-C2f structure is proposed, which replaces the standard convolution and bottleneck module with a GSCov and partial convolution, respectively, to achieve a lighter and faster block. Experiments on three underwater datasets (RUOD, UTDAC2020, and URPC2022) show that the proposed method achieves mAP50 of 86.8%, 84.3%, and 84.7% respectively, with a speed of 156 FPS on NVIDIA A30 GPUs, meeting real-time detection requirements. Compared to YOLOv8s, the model volume is reduced by an average of 24%, and mAP accuracy is enhanced on all three datasets. The paper also discusses related works in underwater object detection and lightweight network designs, highlighting the balance between model size, speed, and accuracy.