<p>Camouflaged object detection endeavors to recognize entities that blend seamlessly with their surroundings. Current methodologies struggle to efficiently integrate global context, local features and boundary details to produce precise predictions. Hence, we introduce HTBNet, a novel triple-branch architecture designed to extract and consolidate diverse information effectively, with the primary goal of enhancing the accurate identification of camouflaged targets. Specifically, the boundary exploration module extracts boundary information in a branch, while the token-enhanced localization module with spatial attention focuses on exact location information from three high-level feature layers in another branch. Subsequently, the boundary and positional information are fed into the triple attention cross fusion Module, employing reverse attention to extract crucial elements from the aforementioned features and facilitate the fusion of multi-scale features. Furthermore, the deep overlapping refinement module integrated a U-shaped residual architecture is incorporated to enhance fusion results, ultimately contributing to meticulous predictions. Experimental results demonstrate that the proposed method achieves state-of-the-art performance across three benchmark datasets. It improves the <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="530_2025_1853_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="26" /> </InlineMediaObject> <EquationSource Format="TEX">\({E_m}\)</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>E</mi> <mi>m</mi> </msub> </math></EquationSource> </InlineEquation> to 0.937 on NC4K and reduces the <i>MAE</i> by 3.0% compared to existing 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

Hierarchical triple-branch network for camouflaged object detection via progressive feature refinement

  • Qing Pan,
  • Zuqing Huang,
  • Nili Tian

摘要

Camouflaged object detection endeavors to recognize entities that blend seamlessly with their surroundings. Current methodologies struggle to efficiently integrate global context, local features and boundary details to produce precise predictions. Hence, we introduce HTBNet, a novel triple-branch architecture designed to extract and consolidate diverse information effectively, with the primary goal of enhancing the accurate identification of camouflaged targets. Specifically, the boundary exploration module extracts boundary information in a branch, while the token-enhanced localization module with spatial attention focuses on exact location information from three high-level feature layers in another branch. Subsequently, the boundary and positional information are fed into the triple attention cross fusion Module, employing reverse attention to extract crucial elements from the aforementioned features and facilitate the fusion of multi-scale features. Furthermore, the deep overlapping refinement module integrated a U-shaped residual architecture is incorporated to enhance fusion results, ultimately contributing to meticulous predictions. Experimental results demonstrate that the proposed method achieves state-of-the-art performance across three benchmark datasets. It improves the \({E_m}\) E m to 0.937 on NC4K and reduces the MAE by 3.0% compared to existing approaches.