Scholarlink

<p>Infrared small target detection faces significant challenges in complex backgrounds and has garnered widespread attention in recent years. To address the trade-off between detection accuracy and computational complexity in current methods, we do not solely pursue higher detection accuracy or IoU values. Instead, we aim to achieve a more lightweight model—one with fewer parameters and reduced computational load—while surpassing the average detection accuracy and IoU values of existing state-of-the-art models. The proposed network introduces a spatial-channel difference convolution module, which efficiently extracts spatial and channel features while maintaining low computational cost. Additionally, considering the characteristics of the SIRST dataset, we design a global feature enhancement Mamba layer that effectively accelerates target localization and ensures precise detection in complex backgrounds. Furthermore, we propose a self-learning joint loss function, which dynamically adjusts weights during training to optimize the contributions of different loss components. The proposed method offers significant advantages in inference speed and parameter size: the FPS is several times higher than that of larger model-based methods, with Params and FLOPs of only 0.024M and 0.168G, respectively, far lower than existing lightweight and full-size models. Its overall performance outperforms current lightweight methods and surpasses some large model-based approaches.</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

LMDNet: a lightweight network for infrared small target detection based on mamba and difference convolution

  • 期刊论文

  • 发表日期：

    2025年9月18日

查看全文
  • Dongyuan Zang,
  • Weihua Su,
  • Zijing Song,
  • Jiabao Huang,
  • Meng Yin,
  • Jun Ma,
  • Shenao Song

摘要

Infrared small target detection faces significant challenges in complex backgrounds and has garnered widespread attention in recent years. To address the trade-off between detection accuracy and computational complexity in current methods, we do not solely pursue higher detection accuracy or IoU values. Instead, we aim to achieve a more lightweight model—one with fewer parameters and reduced computational load—while surpassing the average detection accuracy and IoU values of existing state-of-the-art models. The proposed network introduces a spatial-channel difference convolution module, which efficiently extracts spatial and channel features while maintaining low computational cost. Additionally, considering the characteristics of the SIRST dataset, we design a global feature enhancement Mamba layer that effectively accelerates target localization and ensures precise detection in complex backgrounds. Furthermore, we propose a self-learning joint loss function, which dynamically adjusts weights during training to optimize the contributions of different loss components. The proposed method offers significant advantages in inference speed and parameter size: the FPS is several times higher than that of larger model-based methods, with Params and FLOPs of only 0.024M and 0.168G, respectively, far lower than existing lightweight and full-size models. Its overall performance outperforms current lightweight methods and surpasses some large model-based approaches.