Scholarlink

<p>Conventional few-shot segmentation techniques are constrained by foreground-background binary paradigms, failing to effectively exploit prior knowledge from base classes. While existing generalized few-shot semantic segmentation (GFSS) methods enable cross-category joint segmentation, they suffer performance degradation due to neglected latent semantic correlations between base and novel classes. Therefore, we propose a Contrastive Learning and Orthogonal Decoupling-based GFSS model (CLOD-GFSS) featuring a two-phase training strategy to balance base-class classification and novel-class generalization. Our hierarchical context-aware architecture integrates global semantics with local details through a multi-scale anchor representation system, dynamically optimizing intra-class compactness and inter-class separation via contrastive learning objectives. The hyperspherical orthogonal decoupling mechanism constrains feature space geometry to mitigate base-class feature drift while enhancing discriminative ability for novel classes. Experimental results on <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\text {PASCAL-5}^i\)</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mtext>PASCAL-5</mtext> <mi>i</mi> </msup> </math></EquationSource> </InlineEquation> and <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\text {COCO-20}^i\)</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mtext>COCO-20</mtext> <mi>i</mi> </msup> </math></EquationSource> </InlineEquation> demonstrate that CLOD-GFSS achieves a significant improvement in novel-class generalization while maintaining the ability of base-class classification.</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

Generalized few-shot semantic segmentation via contrastive learning and orthogonal decoupling

  • 期刊论文

  • 发表日期：

    2026年3月9日

查看全文
  • Lulu Jiang,
  • Yaozheng Xia,
  • Shaorong Wang

摘要

Conventional few-shot segmentation techniques are constrained by foreground-background binary paradigms, failing to effectively exploit prior knowledge from base classes. While existing generalized few-shot semantic segmentation (GFSS) methods enable cross-category joint segmentation, they suffer performance degradation due to neglected latent semantic correlations between base and novel classes. Therefore, we propose a Contrastive Learning and Orthogonal Decoupling-based GFSS model (CLOD-GFSS) featuring a two-phase training strategy to balance base-class classification and novel-class generalization. Our hierarchical context-aware architecture integrates global semantics with local details through a multi-scale anchor representation system, dynamically optimizing intra-class compactness and inter-class separation via contrastive learning objectives. The hyperspherical orthogonal decoupling mechanism constrains feature space geometry to mitigate base-class feature drift while enhancing discriminative ability for novel classes. Experimental results on \(\text {PASCAL-5}^i\) PASCAL-5 i and \(\text {COCO-20}^i\) COCO-20 i demonstrate that CLOD-GFSS achieves a significant improvement in novel-class generalization while maintaining the ability of base-class classification.