Scholarlink

<p>Due to the discrepancy between training and detected images from different domains, the generalization of object detection models trained on clear weather dataset is poor for foggy scenes. To address the domain shift problems, a perceptual CycleGAN-differential information YOLO (PC-DIYOLO) model was proposed based on the perceptual Cycle-Consistent Generative Adversarial Network (CycleGAN). A perceptual CycleGAN was designed with perceptual loss function, focusing on the semantic content and visual perception of the images to make them more aligned with human visual perception. Then, a differential information YOLO (DIYOLO) was introduced for cross-domain object detection tasks. And an auxiliary domain was constructed during the transfer process from the source domain to the target domain. On this basis, an adaptive module for instance-level features based on the YOLOv8 network was designed. Finally, the proposed method was trained and evaluated with three levels of fog datasets under Foggy Cityscapes, including Foggy Cityscapes0.005, Foggy Cityscapes0.01, and Foggy Cityscapes0.02. Experimental results indicated that, the proposed algorithm achieved a mAP of 57.50% under foggy weather conditions, which enhanced the object detection performance effectively in foggy 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

Cross-Domain Detection Method for Intelligent Vehicles in Foggy Scenes Based on Perceptual Loss

  • 期刊论文

  • 发表日期：

    2025年9月8日

查看全文
  • Feixiang Xu,
  • Pengcheng Wang,
  • Chen Zhou,
  • He Jiang,
  • Lige Xue,
  • Deqiang Cheng

摘要

Due to the discrepancy between training and detected images from different domains, the generalization of object detection models trained on clear weather dataset is poor for foggy scenes. To address the domain shift problems, a perceptual CycleGAN-differential information YOLO (PC-DIYOLO) model was proposed based on the perceptual Cycle-Consistent Generative Adversarial Network (CycleGAN). A perceptual CycleGAN was designed with perceptual loss function, focusing on the semantic content and visual perception of the images to make them more aligned with human visual perception. Then, a differential information YOLO (DIYOLO) was introduced for cross-domain object detection tasks. And an auxiliary domain was constructed during the transfer process from the source domain to the target domain. On this basis, an adaptive module for instance-level features based on the YOLOv8 network was designed. Finally, the proposed method was trained and evaluated with three levels of fog datasets under Foggy Cityscapes, including Foggy Cityscapes0.005, Foggy Cityscapes0.01, and Foggy Cityscapes0.02. Experimental results indicated that, the proposed algorithm achieved a mAP of 57.50% under foggy weather conditions, which enhanced the object detection performance effectively in foggy weather.