电子测量与仪器学报

朱圣博,魏利胜,高港,郑泊文.基于改进YOLOv8的光学遥感小型船舶检测算法[J].电子测量与仪器学报,2024,38(10):48-57

基于改进YOLOv8的光学遥感小型船舶检测算法

Optical remote sensing small ship detection algorithmbased on improved YOLOv8s

DOI：

中文关键词: 遥感图像目标检测小型船舶特征提取动态检测头 YOLOv8s

英文关键词:remote sensing image object detection small ship feature extraction dynamic detection head YOLOv8s

基金项目:安徽省教育厅自然科学研究重大基金资助项目(KJ2020ZD39)、安徽省高等学校省级质量工程项目(2023cxtd057)资助

作者	单位
朱圣博	安徽工程大学电气工程学院芜湖241000
魏利胜	安徽工程大学电气工程学院芜湖241000
高港	安徽工程大学电气工程学院芜湖241000
郑泊文	安徽工程大学电气工程学院芜湖241000

Author	Institution
Zhu Shengbo	School of Electrical Engineering, Anhui Polytechnic University, Wuhu 241000，China
Wei Lisheng	School of Electrical Engineering, Anhui Polytechnic University, Wuhu 241000，China
Gao Gang	School of Electrical Engineering, Anhui Polytechnic University, Wuhu 241000，China
Zheng Bowen	School of Electrical Engineering, Anhui Polytechnic University, Wuhu 241000，China

摘要点击次数: 484

全文下载次数: 1386

中文摘要:

针对海陆边界、近岸岩礁等复杂海洋场景下,光学遥感小型船舶检测成像特征不明显、目标占比小的问题,提出一种改进YOLOv8s的小型船舶检测方法。首先,在颈部层引入浅层特征图的基础上修改预测层,平衡浅层位置信息和深层语义信息的权重,增强模型对小目标的关注度;其次,采用融合FasterNet Block和高效多尺度注意力机制的C2f-FE模块,利用通道分组和跨通道信息交互,加强对小型船舶的特征提取,并降低模型参数;最后,采用动态检测头模块,在不同预测层级上提高模型对不同空间尺度、任务目标的检测能力。实验结果表明,与原始YOLOV8s模型相比,改进模型的参数量减少42.3%,在MASATI数据集上,改进模型的检测精度mAP50和mAP50:95值分别提高4.2%和2.2%,在DOTA-Ship和DOTA-Small Vehicle数据集上,改进模型的检测精度mAP50:95值分别提高1.7%和1.4%。由此可知,改进模型不仅有效地实现轻量化、高精度的小型船舶检测,而且满足在遥感场景下泛化小目标的高精度检测。

英文摘要:

Aiming at the problem that the imaging features are inconspicuous and the proportion of objects is small in the optical remote sensing small ship detection under the complex marine scenes, such as sea-lean boundary and near-shore rocky reefs, an improved small ship detection method based on YOLOv8s is proposed. Firstly, the prediction layers are modified based on the introduction of shallow feature maps in the neck layers, which balances the weights of shallow locational information and deep semantic information, and enhances the attention of the model to small objects. Secondly, the C2f-FE module is adopted to utilize the channel grouping and the cross-channel information interactions, enhance the feature extraction of small ships, and reduce the model parameters, which merges the FasterNet Block and the efficient multi-scale attention mechanism. Finally, the dynamic detection head module is employed to improve detection capability of the model on different spatial scales and object tasks at different prediction layers. The experimental results show that compared with the original YOLOv8s model, the improved model reduces the number of parameters by 42.3%, the detection accuracy mAP50 and mAP50:95 values are improved by 4.2% and 2.2% on the MASATI dataset, and mAP50:95 values are improved by 1.7% and 1.4% on the DOTA-Ship and DOTA-Small Vehicle datasets, respectively. It can be concluded that the improved model not only achieves lightweight and accurate detection of small ships, but also satisfies the high-accuracy detection for the generalized of small objects in remote sensing scenarios.

查看全文查看/发表评论下载PDF阅读器