Scholarlink

<p>Salient object detection (SOD) is vital for computer vision and involves the key challenge of capturing dense pixel-level information while accurately identifying target details. However, existing methods still struggle with missed detection and the false segmentation of salient objects in complex scenarios, particularly under conditions where resources are limited. To overcome these limitations, this paper introduces a novel Hilbert-Huang transform-guided correlation network (HHTNet) for saliency inference. This architecture achieves significant advances through two key modules. The cross-correlation fusion module (CCFM) achieves accurate cross-sample feature matching and localization by establishing dense correlation maps, thereby enhancing saliency prediction. The Hilbert-Huang transform-guided local attention module (HHT-LAM) combines frequency domain analysis with a spatial channel attention mechanism to perform dual feature calibration. This improves feature discrimination while achieving collaborative enhancement and selective refinement of features. Extensive experiments on five datasets demonstrate that our approach achieves scores of 0.810 (0.044), 0.907 (0.041), 0.895 (0.033), 0.816 (0.073), and 0.751 (0.056) on the <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11760_2025_4760_Article_IEq1.gif" Format="GIF" Height="19" Rendition="HTML" Resolution="72" Type="Linedraw" Width="21" /> </InlineMediaObject> <EquationSource Format="TEX">\(F_{\beta }\)</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>F</mi> <mi>β</mi> </msub> </math></EquationSource> </InlineEquation> (MAE) metric respectively. Furthermore, the proposed approach requires only 3.18M parameters and 4.87G FLOPs, striking a balance between accuracy and efficiency and providing practical possibilities for edge deployment.</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

HHTNet: Hilbert-Huang transform-guided correlation network for salient object detection

  • 期刊论文

  • 发表日期：

    2025年9月11日

查看全文
  • Baoyu Wang,
  • Bo Jiang,
  • Pingping Cao

摘要

Salient object detection (SOD) is vital for computer vision and involves the key challenge of capturing dense pixel-level information while accurately identifying target details. However, existing methods still struggle with missed detection and the false segmentation of salient objects in complex scenarios, particularly under conditions where resources are limited. To overcome these limitations, this paper introduces a novel Hilbert-Huang transform-guided correlation network (HHTNet) for saliency inference. This architecture achieves significant advances through two key modules. The cross-correlation fusion module (CCFM) achieves accurate cross-sample feature matching and localization by establishing dense correlation maps, thereby enhancing saliency prediction. The Hilbert-Huang transform-guided local attention module (HHT-LAM) combines frequency domain analysis with a spatial channel attention mechanism to perform dual feature calibration. This improves feature discrimination while achieving collaborative enhancement and selective refinement of features. Extensive experiments on five datasets demonstrate that our approach achieves scores of 0.810 (0.044), 0.907 (0.041), 0.895 (0.033), 0.816 (0.073), and 0.751 (0.056) on the \(F_{\beta }\) F β (MAE) metric respectively. Furthermore, the proposed approach requires only 3.18M parameters and 4.87G FLOPs, striking a balance between accuracy and efficiency and providing practical possibilities for edge deployment.