Scholarlink

<p>The widespread application of underactuated Unmanned Surface Vehicles (USVs) in ship detection and maritime surveillance has highlighted the need for advanced target detection models. However, hardware limitations and environmental constraints impact the perception capability of USVs. This increases the demands on both the speed and accuracy of lightweight detection algorithms. In this study, we adopted a progressive fusion approach to optimize feature information integration, balancing computational resources and accuracy. The network, named PFMultiNet, is based on Yolov7-tiny. It utilizes an interaction strategy between sub-backbones of varying depths in its backbone to enhance the extraction of positional features. Compared to multi-backbone networks with the same depth, this approach leverages the advantages of a multi-backbone architecture while mitigating the computational burden. Feature Interaction Network (FIN) is proposed, combining multi-scale information and cross-layer attention weights, significantly improving the extraction of contour features. Progressive Fusion Network (PFN) is adopted as the Neck network, where a hierarchical fusion strategy is used to effectively reduce redundant information while preserving the expression of key features. The proposed Progressive Fusion Module (PFM) integrates information from different scales layer by layer, optimizing the feature representation capability and enhancing the model’s adaptability and accuracy in handling multi-scale targets. Experimental results indicate that PFMultiNet outperforms the existing state-of-the-art networks, achieving an accuracy of 97.1% with a model size of only 15.7&#xa0;MB, while maintaining an FPS of 72. Additionally, this paper delves into the optimal combinations of multi-backbone networks, providing valuable insights for research in the field of ship detection.</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

PFMultiNet: a progressive fusion ship detection network for underactuated unmanned surface vehicles in complex environments

  • 期刊论文

  • 发表日期：

    2025年7月3日

查看全文
  • Weina Zhou,
  • Shao Wei,
  • Wenhua Hu

摘要

The widespread application of underactuated Unmanned Surface Vehicles (USVs) in ship detection and maritime surveillance has highlighted the need for advanced target detection models. However, hardware limitations and environmental constraints impact the perception capability of USVs. This increases the demands on both the speed and accuracy of lightweight detection algorithms. In this study, we adopted a progressive fusion approach to optimize feature information integration, balancing computational resources and accuracy. The network, named PFMultiNet, is based on Yolov7-tiny. It utilizes an interaction strategy between sub-backbones of varying depths in its backbone to enhance the extraction of positional features. Compared to multi-backbone networks with the same depth, this approach leverages the advantages of a multi-backbone architecture while mitigating the computational burden. Feature Interaction Network (FIN) is proposed, combining multi-scale information and cross-layer attention weights, significantly improving the extraction of contour features. Progressive Fusion Network (PFN) is adopted as the Neck network, where a hierarchical fusion strategy is used to effectively reduce redundant information while preserving the expression of key features. The proposed Progressive Fusion Module (PFM) integrates information from different scales layer by layer, optimizing the feature representation capability and enhancing the model’s adaptability and accuracy in handling multi-scale targets. Experimental results indicate that PFMultiNet outperforms the existing state-of-the-art networks, achieving an accuracy of 97.1% with a model size of only 15.7 MB, while maintaining an FPS of 72. Additionally, this paper delves into the optimal combinations of multi-backbone networks, providing valuable insights for research in the field of ship detection.