Scholarlink

<p>In autonomous driving, image degradation caused by severe weather (haze, rain, snow) significantly reduces the accuracy and real-time performance of traffic sign recognition, with prominent bottlenecks of insufficient generalization in single-modal methods and high computational costs in multi-modal methods. To address this, this paper proposes the YOLOv8n-MTSA model based on a Multi-Teacher-Single-Student Architecture (MTSA), featuring a core innovative phased lightweight learning mechanism. In the Knowledge Consolidation (KC) stage, Multi-Teacher Collaborative Knowledge Distillation (MT-KD) aggregates single-weather prior knowledge; Cross-Modal Feature Alignment (CFA) unifies features into a shared semantic space, paired with Bidirectional Feature Verification (BFV) to avoid detail distortion, and Soft Label Contrastive Learning (SCL) reduces multi-modal learning difficulty. In the Knowledge Evaluation (KE) stage, Hard Label Contrastive Learning (HCL) enhances real-scene discriminability, combined with pixel-level loss to optimize restoration. Experiments on a mixed-weather dataset show the model achieves image restoration with PSNR=34.023 and SSIM=0.927, detection mAP@50 of 88.2% (4.1% higher than YOLOv8n baseline), and maintains 70.4 FPS (14.2 ms per image) real-time performance, providing an efficient solution for in-vehicle deployment under complex weather.</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

Enhanced Traffic Sign Detection in Severe Weather: A Multi-Teacher-Single-Student Knowledge Distillation Approach

  • 期刊论文

  • 发表日期：

    2026年3月31日

查看全文
  • Xuan Sun,
  • Yancan Wu,
  • Lingyun Zhang

摘要

In autonomous driving, image degradation caused by severe weather (haze, rain, snow) significantly reduces the accuracy and real-time performance of traffic sign recognition, with prominent bottlenecks of insufficient generalization in single-modal methods and high computational costs in multi-modal methods. To address this, this paper proposes the YOLOv8n-MTSA model based on a Multi-Teacher-Single-Student Architecture (MTSA), featuring a core innovative phased lightweight learning mechanism. In the Knowledge Consolidation (KC) stage, Multi-Teacher Collaborative Knowledge Distillation (MT-KD) aggregates single-weather prior knowledge; Cross-Modal Feature Alignment (CFA) unifies features into a shared semantic space, paired with Bidirectional Feature Verification (BFV) to avoid detail distortion, and Soft Label Contrastive Learning (SCL) reduces multi-modal learning difficulty. In the Knowledge Evaluation (KE) stage, Hard Label Contrastive Learning (HCL) enhances real-scene discriminability, combined with pixel-level loss to optimize restoration. Experiments on a mixed-weather dataset show the model achieves image restoration with PSNR=34.023 and SSIM=0.927, detection mAP@50 of 88.2% (4.1% higher than YOLOv8n baseline), and maintains 70.4 FPS (14.2 ms per image) real-time performance, providing an efficient solution for in-vehicle deployment under complex weather.