Scholarlink

<p>The core challenge of camouflaged object detection (COD) lies in identifying concealed objects that blend deeply with their background and exhibit extremely blurred boundaries. While existing approaches attempt to enhance feature learning by integrating the segment anything model (SAM) and introducing static edge prompts, their reliance on fixed prompts limits the efficient utilization of edge information under complex boundaries. To address this, we propose SAM2-DEGNet, a two-stage edge-guided network based on the SAM2 Hiera backbone. Unlike previous COD models primarily relying on SAM prompting mechanisms, our approach synergistically leverages two complementary stages: During the dynamic edge optimization stage, edge priors guide adaptive dynamic edge enhancement of intermediate results, integrated with edge confidence scores from upper-level predictions. In the edge context reconstruction stage, we further introduce a combination of fourth-order Runge–Kutta and convolutional operators to reconstruct edges without relying on edge priors, effectively correcting potential error guidance from the first stage. This complementary interaction overcomes the limitations of static prompting methods while significantly enhancing the model’s adaptability to complex camouflaged boundaries. Experimental results demonstrate that SAM2-DEGNet outperforms existing state-of-the-art methods across three common COD datasets. Notably, it achieves an average performance gain of up to 7.6% on the highly camouflaged CAMO dataset, validating the proposed method’s strong adaptability and robustness in complex camouflage scenarios. The code will be available at <a href="https://github.com/zzp1314525/SAM2-DEGNet">https://github.com/zzp1314525/SAM2-DEGNet</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

SAM2-DEGNet: dual-stage edge guidance network for camouflaged object detection using SAM2

  • 期刊论文

  • 发表日期：

    2025年12月13日

查看全文
  • Zhenpeng Zhong,
  • Ge Jiao,
  • Guangcheng Li,
  • Tanghui Liu,
  • Yang Liu

摘要

The core challenge of camouflaged object detection (COD) lies in identifying concealed objects that blend deeply with their background and exhibit extremely blurred boundaries. While existing approaches attempt to enhance feature learning by integrating the segment anything model (SAM) and introducing static edge prompts, their reliance on fixed prompts limits the efficient utilization of edge information under complex boundaries. To address this, we propose SAM2-DEGNet, a two-stage edge-guided network based on the SAM2 Hiera backbone. Unlike previous COD models primarily relying on SAM prompting mechanisms, our approach synergistically leverages two complementary stages: During the dynamic edge optimization stage, edge priors guide adaptive dynamic edge enhancement of intermediate results, integrated with edge confidence scores from upper-level predictions. In the edge context reconstruction stage, we further introduce a combination of fourth-order Runge–Kutta and convolutional operators to reconstruct edges without relying on edge priors, effectively correcting potential error guidance from the first stage. This complementary interaction overcomes the limitations of static prompting methods while significantly enhancing the model’s adaptability to complex camouflaged boundaries. Experimental results demonstrate that SAM2-DEGNet outperforms existing state-of-the-art methods across three common COD datasets. Notably, it achieves an average performance gain of up to 7.6% on the highly camouflaged CAMO dataset, validating the proposed method’s strong adaptability and robustness in complex camouflage scenarios. The code will be available at https://github.com/zzp1314525/SAM2-DEGNet.