This paper proposes an enhanced YOLOv8 network for efficient small-object detection in underwater images. The method improves detection accuracy while reducing model size and computational complexity. The backbone of YOLOv8 is replaced with FasterNet-T0, reducing model parameters by 22.52%, FLOPS by 23.59%, and model size by 22.73%. A prediction head for small objects is added, increasing the number of channels in high-resolution feature maps and decreasing those in low-resolution maps, leading to a 1.2% improvement in small-object detection accuracy. Deformable ConvNets and Coordinate Attention are used in the neck part to enhance detection of irregularly shaped and densely occluded small targets. The model achieves 52.12% AP on the UTDAC2020 dataset, surpassing the performance of the larger YOLOv8l model. When input resolution is increased to 1280×1280 pixels, the AP reaches 53.18%, making it the state-of-the-art model for the UTDAC2020 dataset. The model also achieves 84.4% mAP on the Pascal VOC dataset, with significantly reduced model parameters compared to previous detectors. The experimental results demonstrate the effectiveness of the proposed lightweight method in underwater environments and its generalization to common datasets. The method achieves high accuracy with lightweight parameters (8.5 M), low FLOPS (25.5 B), and a small model size (17 MB), while maintaining an AP of 53.18% and AP50 of 86.21%. The results show that the method outperforms existing detectors in terms of lightweighting and model accuracy. The method is applicable to various underwater marine scenarios and demonstrates strong performance on unseen datasets. The paper concludes that the enhanced YOLOv8 model achieves state-of-the-art performance on the UTDAC2020 dataset, maintaining lightweight parameters while achieving high accuracy. Future work will focus on further improving the balance between model size and detection accuracy in underwater object detection.This paper proposes an enhanced YOLOv8 network for efficient small-object detection in underwater images. The method improves detection accuracy while reducing model size and computational complexity. The backbone of YOLOv8 is replaced with FasterNet-T0, reducing model parameters by 22.52%, FLOPS by 23.59%, and model size by 22.73%. A prediction head for small objects is added, increasing the number of channels in high-resolution feature maps and decreasing those in low-resolution maps, leading to a 1.2% improvement in small-object detection accuracy. Deformable ConvNets and Coordinate Attention are used in the neck part to enhance detection of irregularly shaped and densely occluded small targets. The model achieves 52.12% AP on the UTDAC2020 dataset, surpassing the performance of the larger YOLOv8l model. When input resolution is increased to 1280×1280 pixels, the AP reaches 53.18%, making it the state-of-the-art model for the UTDAC2020 dataset. The model also achieves 84.4% mAP on the Pascal VOC dataset, with significantly reduced model parameters compared to previous detectors. The experimental results demonstrate the effectiveness of the proposed lightweight method in underwater environments and its generalization to common datasets. The method achieves high accuracy with lightweight parameters (8.5 M), low FLOPS (25.5 B), and a small model size (17 MB), while maintaining an AP of 53.18% and AP50 of 86.21%. The results show that the method outperforms existing detectors in terms of lightweighting and model accuracy. The method is applicable to various underwater marine scenarios and demonstrates strong performance on unseen datasets. The paper concludes that the enhanced YOLOv8 model achieves state-of-the-art performance on the UTDAC2020 dataset, maintaining lightweight parameters while achieving high accuracy. Future work will focus on further improving the balance between model size and detection accuracy in underwater object detection.