Scholarlink

<p>Camouflaged object detection (COD) aims to identify objects seamlessly embedded in their surroundings. Current COD methods often face challenges due to their large model parameters and extensive computational requirements, limiting their practicality. To address these issues, this paper proposes a lightweight RGB-D camouflaged object detection framework that leverages hybrid attention mechanisms and boundary perception. By adopting MobileNetV2 and standard convolutions as the backbone network, our model maintains a lightweight design while improving feature extraction through boundary boosting and transfer learning techniques. Additionally, spatial and semantic information is enriched using a spatial enhancement module and multiple attention modules. A boundary perception module is embedded into each side output to effectively capture region and boundary features, enhancing detection accuracy, especially around boundaries. Experimental results demonstrate that our model achieves a competitive performance compared to state-of-the-art methods, with a model size of only 16.41MB and a high frame rate of 108 FPS on a single NVIDIA 3090 GPU. This lightweight design ensures our model’s applicability for mobile and edge devices. The code will be available at <a href="https://github.com/YangLiu353/LHA.git">https://github.com/YangLiu353/LHA.git</a>.</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

Lightweight hybrid attention RGB-D networks for accurate camouflaged object detection

  • 期刊论文

  • 发表日期：

    2025年3月18日

查看全文
  • Yang Liu,
  • Shuhan Chen,
  • Haonan Tang,
  • Shiyu Wang

摘要

Camouflaged object detection (COD) aims to identify objects seamlessly embedded in their surroundings. Current COD methods often face challenges due to their large model parameters and extensive computational requirements, limiting their practicality. To address these issues, this paper proposes a lightweight RGB-D camouflaged object detection framework that leverages hybrid attention mechanisms and boundary perception. By adopting MobileNetV2 and standard convolutions as the backbone network, our model maintains a lightweight design while improving feature extraction through boundary boosting and transfer learning techniques. Additionally, spatial and semantic information is enriched using a spatial enhancement module and multiple attention modules. A boundary perception module is embedded into each side output to effectively capture region and boundary features, enhancing detection accuracy, especially around boundaries. Experimental results demonstrate that our model achieves a competitive performance compared to state-of-the-art methods, with a model size of only 16.41MB and a high frame rate of 108 FPS on a single NVIDIA 3090 GPU. This lightweight design ensures our model’s applicability for mobile and edge devices. The code will be available at https://github.com/YangLiu353/LHA.git.