Scholarlink

<p>Traffic sign detection under foggy conditions poses critical challenges for autonomous driving due to atmospheric scattering that degrades image contrast and blurs edge information. Existing Transformer-based detectors lack explicit mechanisms to handle fog-induced degradation. We propose CLEAR-DETR, a fog-robust framework integrating three complementary components: (1)&#xa0;CSP-EMFI backbone with EdgeEnhancer modules for frequency-domain edge restoration; (2)&#xa0;MSFFN neck with spatial-depth transformation and hybrid attention for multi-scale fusion; and (3)&#xa0;AIFI-RepBN framework replacing LayerNorm with re-parameterized batch normalization for efficient inference. We construct Foggy-100K, comprising 8,581 images across diverse fog densities, for systematic evaluation. On Foggy-100K, CLEAR-DETR achieves 87.1% <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\hbox {mAP}_{50}\)</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mtext>mAP</mtext> <mn>50</mn> </msub> </math></EquationSource> </InlineEquation> with 15.9M parameters (20.1% reduction vs. RT-DETR-R18’s 19.9M, 83.7%), improving accuracy by 3.4 percentage points. Ablation studies confirm complementary module gains with partial functional overlap, with MSFFN contributing the largest individual improvement (+2.1 <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\hbox {mAP}_{50}\)</EquationSource> <EquationSource Format="MATHML"><math> <msub> <mtext>mAP</mtext> <mn>50</mn> </msub> </math></EquationSource> </InlineEquation>).</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

CLEAR-DETR: Cross-weather Low-visibility Enhanced Atmospheric Recognition Transformer for Robust Traffic Sign Detection

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Xiangyu Kong,
  • Xinying Chen

摘要

Traffic sign detection under foggy conditions poses critical challenges for autonomous driving due to atmospheric scattering that degrades image contrast and blurs edge information. Existing Transformer-based detectors lack explicit mechanisms to handle fog-induced degradation. We propose CLEAR-DETR, a fog-robust framework integrating three complementary components: (1) CSP-EMFI backbone with EdgeEnhancer modules for frequency-domain edge restoration; (2) MSFFN neck with spatial-depth transformation and hybrid attention for multi-scale fusion; and (3) AIFI-RepBN framework replacing LayerNorm with re-parameterized batch normalization for efficient inference. We construct Foggy-100K, comprising 8,581 images across diverse fog densities, for systematic evaluation. On Foggy-100K, CLEAR-DETR achieves 87.1% \(\hbox {mAP}_{50}\) mAP 50 with 15.9M parameters (20.1% reduction vs. RT-DETR-R18’s 19.9M, 83.7%), improving accuracy by 3.4 percentage points. Ablation studies confirm complementary module gains with partial functional overlap, with MSFFN contributing the largest individual improvement (+2.1 \(\hbox {mAP}_{50}\) mAP 50 ).