Scholarlink

<p>Accurate ship detection in complex maritime environments remains a challenging task due to illumination variations, occlusions, dense distributions, and small target sizes. To address these challenges, this paper proposes KAN-YOLO, a novel ship detection framework that integrates nonlinear modeling and global feature extraction within the YOLO architecture. Specifically, a KAN-enhanced depthwise separable convolution module is proposed to enhance the model’s nonlinear feature representation, enabling more effective discrimination of complex object boundaries and subtle visual cues. Furthermore, a KAN-enhanced state space duality module is designed to capture long-range dependencies and global contextual information with linear computational complexity. Finally, a KAN-Head is developed to further improve localization precision under complex maritime scenarios. Extensive experiments on the SeaShips and FAIR1M1.0 datasets demonstrate that the proposed method KAN-YOLO achieves superior accuracy and robustness compared with existing YOLO variants, attaining the highest mAP@50:95 of 82.7% and 43.0%, respectively, while maintaining lightweight complexity with only 2.55M parameters and 5.9 GFLOPs, making it well-suited for maritime surveillance and intelligent waterway transportation applications.</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

KAN-YOLO: a nonlinear-enhanced and state space duality-driven YOLO network for robust ship detection in complex maritime scenes

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Jingzhu Li,
  • Lintao Xu,
  • Jianbin Jiao,
  • Qimeng Huang,
  • Shunqing Yang,
  • Pupu Wang,
  • Ao Jiao

摘要

Accurate ship detection in complex maritime environments remains a challenging task due to illumination variations, occlusions, dense distributions, and small target sizes. To address these challenges, this paper proposes KAN-YOLO, a novel ship detection framework that integrates nonlinear modeling and global feature extraction within the YOLO architecture. Specifically, a KAN-enhanced depthwise separable convolution module is proposed to enhance the model’s nonlinear feature representation, enabling more effective discrimination of complex object boundaries and subtle visual cues. Furthermore, a KAN-enhanced state space duality module is designed to capture long-range dependencies and global contextual information with linear computational complexity. Finally, a KAN-Head is developed to further improve localization precision under complex maritime scenarios. Extensive experiments on the SeaShips and FAIR1M1.0 datasets demonstrate that the proposed method KAN-YOLO achieves superior accuracy and robustness compared with existing YOLO variants, attaining the highest mAP@50:95 of 82.7% and 43.0%, respectively, while maintaining lightweight complexity with only 2.55M parameters and 5.9 GFLOPs, making it well-suited for maritime surveillance and intelligent waterway transportation applications.