Scholarlink

<p>With the diversification of data collection methods, the generalization ability of object detection algorithms is challenged. Therefore, this paper proposes a marine ship target detection algorithm based on the improved YOLOv10. Firstly, this paper proposes the Three-stage Concat Fusion Arithmetic (TCFusion) algorithm to improve the feature loss problem of the traditional feature fusion module and fully retain the semantic information of some upper-level feature maps and the position information of lower-level feature maps. Secondly, in view of the inevitable feature loss caused by pooling operations and step-convolution operations, this paper proposes a dual-path simulation large-kernel convolution algorithm (DSLC), which retains the unique ability of large kernel convolution to perceive spatial position information without increasing the computational load and reducing feature loss. Finally, in this paper, an Enhanced Adaptive Convolutional Network (EAC) is designed to enable the algorithm to have adaptive target localization ability. The algorithm proposed in this paper is applied in HRSID (SAR), LEVIR (Remote sensing), and RCVN (Infrared). The detection accuracies reached 95.3%, 84.5%, and 79.1% respectively on three different types of datasets. Compared with the mainstream algorithm, they were improved by 1.2%, 1.7%, and 0.8% respectively.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

MPEANet: a strong generalisation capability maritime vessel target detection method

  • 期刊论文

  • 发表日期：

    2025年10月18日

查看全文
  • Pengfei He,
  • Chunhao Bo,
  • Yong Huang,
  • Xia Liu,
  • Guoxing Li,
  • Ningbo Liu

摘要

With the diversification of data collection methods, the generalization ability of object detection algorithms is challenged. Therefore, this paper proposes a marine ship target detection algorithm based on the improved YOLOv10. Firstly, this paper proposes the Three-stage Concat Fusion Arithmetic (TCFusion) algorithm to improve the feature loss problem of the traditional feature fusion module and fully retain the semantic information of some upper-level feature maps and the position information of lower-level feature maps. Secondly, in view of the inevitable feature loss caused by pooling operations and step-convolution operations, this paper proposes a dual-path simulation large-kernel convolution algorithm (DSLC), which retains the unique ability of large kernel convolution to perceive spatial position information without increasing the computational load and reducing feature loss. Finally, in this paper, an Enhanced Adaptive Convolutional Network (EAC) is designed to enable the algorithm to have adaptive target localization ability. The algorithm proposed in this paper is applied in HRSID (SAR), LEVIR (Remote sensing), and RCVN (Infrared). The detection accuracies reached 95.3%, 84.5%, and 79.1% respectively on three different types of datasets. Compared with the mainstream algorithm, they were improved by 1.2%, 1.7%, and 0.8% respectively.