Scholarlink

<p>Improving the speed and accuracy of ship target detection in maritime environments is crucial for the application of intelligent ship navigation technologies. Complex oceanic conditions, such as waves, lighting, and overlapping multiple targets, often pose challenges, particularly in detecting small ships. To address this, we propose a target detection method based on a Weighted Bidirectional and Deconvolution Feature Pyramid Network (YOLOX-WBDeconv-D). This method enhances the model’s representation ability by selecting specific layers of the feature fusion network and employing cross-scale bidirectional weighted adaptive learning to fully integrate information from both deep and shallow feature maps. Additionally, a deconvolution-based upsampling technique is utilized to strengthen the semantic information of intermediate feature maps, significantly improving the detection efficiency of small targets. Compared to YOLOv7 and YOLOv8, which have been widely verified their maturity and stability in the field of small target detection, the proposed network exhibits more excellent performance in complex marine scenarios. The experimental results show that the YOLOX-WBDeconv-D model achieves similar detection accuracy to YOLOv7 while using only 35.5% of the parameters, and it outperforms YOLOv8 by 3.89% in recall rate while maintaining almost same parameter count.</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

Ship target detection via cross-scale weighted bidirectional learning and deconvolution pyramid

  • 期刊论文

  • 发表日期：

    2025年7月30日

查看全文
  • Weina Zhou,
  • Liangtao Yuan,
  • Jiaming Li,
  • Huafeng Wu

摘要

Improving the speed and accuracy of ship target detection in maritime environments is crucial for the application of intelligent ship navigation technologies. Complex oceanic conditions, such as waves, lighting, and overlapping multiple targets, often pose challenges, particularly in detecting small ships. To address this, we propose a target detection method based on a Weighted Bidirectional and Deconvolution Feature Pyramid Network (YOLOX-WBDeconv-D). This method enhances the model’s representation ability by selecting specific layers of the feature fusion network and employing cross-scale bidirectional weighted adaptive learning to fully integrate information from both deep and shallow feature maps. Additionally, a deconvolution-based upsampling technique is utilized to strengthen the semantic information of intermediate feature maps, significantly improving the detection efficiency of small targets. Compared to YOLOv7 and YOLOv8, which have been widely verified their maturity and stability in the field of small target detection, the proposed network exhibits more excellent performance in complex marine scenarios. The experimental results show that the YOLOX-WBDeconv-D model achieves similar detection accuracy to YOLOv7 while using only 35.5% of the parameters, and it outperforms YOLOv8 by 3.89% in recall rate while maintaining almost same parameter count.