| 摘要: |
| 海洋环境中的目标检测由于成像设备距离较远、海上环境复杂、计算资源匮乏等原因导致图像分辨率低，模型的定位精度差。为此，提出了一种基于改进YOLOv8s的海上小目标检测算法YOLOv8s-Star，旨在通过多维度网络结构优化，不显著增加计算开销和模型复杂度的前提下提高小目标检测精度。引入C2f-Star模块替换原有骨干网络中的C2f模块，增强小目标与背景的对比度，有效解决特征丢失问题。引入反向残差移动模块（Inverted Residual Mobile Block，iRMB）对原有快速-空间金字塔池化层进行改进，通过结合轻量级卷积与自注意力机制以增强模型对海上小目标的细节特征关注度。使用Haar小波变换（Haar Wavelet Downsampling，HWD）对融合特征进行下采样操作，提高模型对边缘细节信息的感知能力，进而实现对海上小目标的精确检测。在自建数据集上的实验结果表明，与基线模型相比，该算法mAP@0.50提高2.6%，模型参数量降低了7.9%。 |
| 关键词: 海面目标 小目标检测 模型轻量化 Haar小波下采样 反向残差移动块 |
| DOI：10.20079/j.issn.1001-893x.241220005 |
|
| 基金项目:国家自然科学基金资助项目（62371085） |
|
| A Sea-surface Small Target Detection Algorithm Based on Improved YOLOv8s |
| SUN Rui,LIN Bin,MA Wenxuan |
| (College of Information Science and Technology，Dalian Maritime University，Dalian 116026，China) |
| Abstract: |
| Target detection in the marine environment is characterized by low image resolution and poor localization accuracy of the model due to the long distance of the imaging equipment，the complexity of the marine environment，and the lack of computational resources.To address above problems，a maritime small target detection algorithm YOLOv8s-Star based on improved YOLOv8s is proposed，aiming to improve the small target detection accuracy without significantly increasing the computational overhead and model complexity.The C2f-Star module is introduced to replace the C2f module in the original backbone network，which enhances the contrast between the small targets and the background，and effectively solves the feature loss problem.The Inverted Residual Mobile Block(iRMB) is introduced to improve the original fast-space pyramid pooling layer，and the model is enhanced to detect the small targets at sea by combining the light-weight convolution and the self-attention mechanism to enhance the model’s attention to the detailed features of small targets at sea.The Haar Wavelet Downsampling(HWD) is used to downsample the fused features to improve the model’s ability to perceive the edge detail information，and then realize the accurate detection of small targets at sea.The experimental results on the self-built dataset show that mAP@050 is improved by 26%，and the number of model parameters is reduced by 7.9%，compared with those of baseline models. |
| Key words: small target detection sea-surface target model lightening Haar wavelet downsampling inverted residual mobile block |
《电讯技术》编辑部 