Scholarlink

<p>Sclera-based biometrics has emerged as a promising modality for identity verification, particularly in unconstrained and mobile settings. The unique and stable vascular patterns of the sclera offer resilience to spoofing and enable recognition under off-angle or partially occluded views. However, accurate sclera segmentation remains challenging due to low contrast with surrounding skin, specular reflections, occlusions (e.g., eyelashes, spectacles), and variable illumination. To address these challenges, benchmarking efforts such as the Sclera Segmentation Benchmarking Competition (SSBC) provide standardized datasets (e.g., MASD, SMD, MOBIUS) for fair comparison. Motivated by these constraints and the need for generalizable models, we propose <b>SwinDANet</b>–a segmentation architecture that integrates hierarchical vision transformers (Swin Transformer) with a densely connected convolutional decoder, enhanced by a Concurrent Spatial–Channel Squeeze-and-Excitation (CSSE) block. This hybrid design combines long-range contextual modeling, local texture representation, and adaptive attention recalibration to improve boundary precision under challenging conditions. In the SSBC-2025 <i>validation</i> evaluation, SwinDANet ranked first in the Synthetic Track by <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(F_1\)</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mi>F</mi> <mn>1</mn> </msub> </math></EquationSource> </InlineEquation> (0.798) and IoU (0.680); its AUC was competitive (0.870) but not the highest. In the Mixed Track (synthetic + real), it placed fourth with <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(F_1{=}0.822\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <msub> <mi>F</mi> <mn>1</mn> </msub> <mo>=</mo> <mn>0.822</mn> </mrow> </math></EquationSource> </InlineEquation>, with gaps of 0.017 to rank-1, 0.016 to rank-2, and 0.004 to rank-3. These validation-based results indicate strong performance on synthetic data and competitive cross-domain generalization; blinded test-set conclusions await release of official test annotations.</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

SwinDANet: leveraging swin transformers with Context-Aware Attention for precise sclera segmentation

  • 期刊论文

  • 发表日期：

    2026年1月9日

查看全文
  • Sabari Nathan,
  • K Uma,
  • Swarna Sethu

摘要

Sclera-based biometrics has emerged as a promising modality for identity verification, particularly in unconstrained and mobile settings. The unique and stable vascular patterns of the sclera offer resilience to spoofing and enable recognition under off-angle or partially occluded views. However, accurate sclera segmentation remains challenging due to low contrast with surrounding skin, specular reflections, occlusions (e.g., eyelashes, spectacles), and variable illumination. To address these challenges, benchmarking efforts such as the Sclera Segmentation Benchmarking Competition (SSBC) provide standardized datasets (e.g., MASD, SMD, MOBIUS) for fair comparison. Motivated by these constraints and the need for generalizable models, we propose SwinDANet–a segmentation architecture that integrates hierarchical vision transformers (Swin Transformer) with a densely connected convolutional decoder, enhanced by a Concurrent Spatial–Channel Squeeze-and-Excitation (CSSE) block. This hybrid design combines long-range contextual modeling, local texture representation, and adaptive attention recalibration to improve boundary precision under challenging conditions. In the SSBC-2025 validation evaluation, SwinDANet ranked first in the Synthetic Track by \(F_1\) F 1 (0.798) and IoU (0.680); its AUC was competitive (0.870) but not the highest. In the Mixed Track (synthetic + real), it placed fourth with \(F_1{=}0.822\) F 1 = 0.822 , with gaps of 0.017 to rank-1, 0.016 to rank-2, and 0.004 to rank-3. These validation-based results indicate strong performance on synthetic data and competitive cross-domain generalization; blinded test-set conclusions await release of official test annotations.