Scholarlink

<p>Rainy nighttime conditions pose severe challenges for camera-based object detection because low illumination, rain streaks, raindrops, blur, and partial occlusion jointly weaken object-defining visual cues. Current detection models are typically optimized for executing in favorable conditions, resulting in incomplete detection performance in adverse environments. In addition, existing restoration-before-detection pipelines often introduce domain shift and increase inference cost, limiting their suitability for real-time visual perception. In this study, we introduce DiffKA, a diffusion-guided knowledge absorption framework that transfers restoration-informed clean features to an object detector during training while retaining only the detector branch during inference. To accomplish this, the DiffKA comprises three subnetworks: a feature improvement subnetwork for low-light enhancement and rain degradation recovery, a knowledge transmission subnetwork for diffusion-guided feature absorption, and an object detection subnetwork for final localization and classification. Comprehensive evaluations on the RNT and RID datasets demonstrate that DiffKA helps the YOLOv10 and YOLOv7 baselines improve up to 4.53% and 7.13%, respectively, and outperforms restoration-based detection pipelines while preserving the inference speed of the baseline detector. The dataset and experimental resources are available at: <a href="https://github.com/val-utehy/DiffKA">https://github.com/val-utehy/DiffKA</a>.</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

Diffusion-guided knowledge absorption for robust object detection under rainy night conditions

  • 期刊论文

  • 发表日期：

    2026年6月29日

查看全文
  • Quoc-Viet Hoang,
  • Trung-Hieu Le

摘要

Rainy nighttime conditions pose severe challenges for camera-based object detection because low illumination, rain streaks, raindrops, blur, and partial occlusion jointly weaken object-defining visual cues. Current detection models are typically optimized for executing in favorable conditions, resulting in incomplete detection performance in adverse environments. In addition, existing restoration-before-detection pipelines often introduce domain shift and increase inference cost, limiting their suitability for real-time visual perception. In this study, we introduce DiffKA, a diffusion-guided knowledge absorption framework that transfers restoration-informed clean features to an object detector during training while retaining only the detector branch during inference. To accomplish this, the DiffKA comprises three subnetworks: a feature improvement subnetwork for low-light enhancement and rain degradation recovery, a knowledge transmission subnetwork for diffusion-guided feature absorption, and an object detection subnetwork for final localization and classification. Comprehensive evaluations on the RNT and RID datasets demonstrate that DiffKA helps the YOLOv10 and YOLOv7 baselines improve up to 4.53% and 7.13%, respectively, and outperforms restoration-based detection pipelines while preserving the inference speed of the baseline detector. The dataset and experimental resources are available at: https://github.com/val-utehy/DiffKA.