Scholarlink

Deteriorating weather, such as precipitation, fog, snow, and sandstorms, greatly debilitates the effectiveness of object detection systems by generating visual noise, reducing contrast, and producing occlusion, thereby posing significant challenges to autonomous driving and surveillance. Conventional methods like DETR, Faster R-CNN, and RCNN register poor robustness in these cases, with mean Average Precision (mAP) ranging from 0.47 to 0.53 under varied weather conditions along with unstable precision and recall values. These limitations are a result of their inability to properly handle weather distortions and domain adaptation between the training set and actual data. To address these challenges, we introduce a hybrid deep learning pipeline that involves DehazeFormer (a transformer-based dehazing network), Diffusion GAN (a domain-adaptive generative adversarial network), and DETR (an end-to-end detection transformer). Our method processes degraded images progressively by first restoring visibility with DehazeFormer, then adapting the enhanced images to a weather-agnostic domain via Diffusion GAN, and finally detecting objects using DETR’s set prediction mechanism. In-depth examination of the DAWN dataset finds our approach far superior to existing methods, registering mAP ratings of 0.6504–0.6521, and precision (0.94–0.95) and recall (0.94–0.95) in sandy, rainy, foggy, and snowy settings. DETR, Faster R-CNN, and RCNN have lower accuracy and consistency, highlighting our hybrid structure. Our model corrects weather distortion, lowers erroneous detections, and is reliable, making it a good choice for safety-critical application in difficult conditions. Future research will extend real-time optimization to additional low-visibility environments.

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

Overcoming Weather Degradation in Object Detection: A Multi-stage Deep Learning Approach with Dehaze Former and Diffusion GAN

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Sandhya Sharma,
  • Arjit Tomar,
  • Pramod Kumar Sagar

摘要

Deteriorating weather, such as precipitation, fog, snow, and sandstorms, greatly debilitates the effectiveness of object detection systems by generating visual noise, reducing contrast, and producing occlusion, thereby posing significant challenges to autonomous driving and surveillance. Conventional methods like DETR, Faster R-CNN, and RCNN register poor robustness in these cases, with mean Average Precision (mAP) ranging from 0.47 to 0.53 under varied weather conditions along with unstable precision and recall values. These limitations are a result of their inability to properly handle weather distortions and domain adaptation between the training set and actual data. To address these challenges, we introduce a hybrid deep learning pipeline that involves DehazeFormer (a transformer-based dehazing network), Diffusion GAN (a domain-adaptive generative adversarial network), and DETR (an end-to-end detection transformer). Our method processes degraded images progressively by first restoring visibility with DehazeFormer, then adapting the enhanced images to a weather-agnostic domain via Diffusion GAN, and finally detecting objects using DETR’s set prediction mechanism. In-depth examination of the DAWN dataset finds our approach far superior to existing methods, registering mAP ratings of 0.6504–0.6521, and precision (0.94–0.95) and recall (0.94–0.95) in sandy, rainy, foggy, and snowy settings. DETR, Faster R-CNN, and RCNN have lower accuracy and consistency, highlighting our hybrid structure. Our model corrects weather distortion, lowers erroneous detections, and is reliable, making it a good choice for safety-critical application in difficult conditions. Future research will extend real-time optimization to additional low-visibility environments.