面向小型船舶遥感智能识别的改进型YOLOv5s模型

An Improved YOLOv5s Model for Intelligent Recognition of Small Ships with Remote Sensing

船舶探测在领土安全和海洋环境保护方面具有重要意义。然而,在复杂环境中准确地检测识别小型船舶仍是存在一定技术难度与挑战。考虑到小型船舶在遥感图像上的特征信息较少,本研究提出了一种基于YOLOv5s的小型船舶检测模型。该模型首先在顶端添加了浅层特征图的检测层,并在中部采用了跳过连接,用以提高小型船只的检测精度;随后提出了一种新颖有效的混合空间金字塔池化(Hybrid Spatial Pyramid Pooling, HSPP)技术,以融合特征图的局部和全局信息;此外,在骨干中采用坐标注意(Coordinate Attention, CA)机制来增强小型船舶的表征,并使用高效IOU (EIOU)作为边界框回归的损失函数,以提高所提模型的定位精度;最终,使用K-means++算法来获得更合理的小型船舶检测锚点。实验选取包括10,133幅图像的多尺度船舶数据集OSSD和公开船舶数据集LEVIR-Ship,结合同类既有模型开展了比较分析,验证模型和算法的有效性。

Ship detection is of great importance in territorial security and marine environmental protection. However, the accurate detection of small ships is a challenging task in complex environments mainly due to small ships having few features on remote sensing images. In this paper, we propose a small ship detection model based on YOLOv5s. Firstly, a detection layer is added with a shallow feature map in the head and skip connections are employed in the neck to improve the detection accuracy of small ships. Then, a novel and effective hybrid spatial pyramid pooling (Hybrid Spatial Pyramid Pooling, HSPP) is proposed to fuse the local and global information of feature maps. In addition, a coordinate attention (Coordinate Attention, CA) mechanism is employed in the backbone to augment the representations of small ships, and efficient IOU (EIOU) is used as the loss function for bounding box regression to enhance the localization accuracy of the proposed model. Finally, K-means++ algorithm is used to obtain more reasonable anchors for small ship detection. We introduce the multiscale ship dataset OSSD that contains 10,133 images, and LEVIR-Ship opening datasets. Experiments on LEVIR-Ship and OSSD validated the effectiveness of our proposed model.