PulseExploreJournal ClubDebatesTrendingResearchersJournals
Instagram
HomeExploreJournal ClubTrending
Synapse
⌘+K
Synapse
March 13, 2026Journal of Image and Graphics0 citationsOpen Access

基于窗口移位注意力的两阶段掩码引导汽车表面高光去除网络

WJWang JiachengZSZhou ShuboPFPan FengSoutheast University

Key Points

  • The aim is to develop a robust method for removing specular highlights from automotive images in real-world scenarios.
  • Constructed a comprehensive automotive highlight dataset under various lighting conditions.
  • Developed a two-stage network architecture with a lightweight U-Net for highlight detection.
  • Implemented a Mask Guided Transformer Unet utilizing shifted window attention for highlight removal.
  • Conducted end-to-end joint training for parameter optimization between detection and removal networks.
  • Achieved a PSNR of 26.37dB on the PSD dataset, ranking second among advanced methods.
  • Improved PSNR to 35.24dB on the SHIQ dataset, exceeding the second place by 1.11dB.
  • Obtained a PSNR of 36.54dB on the SSHR dataset, leading by 0.38dB.
  • Surpassed the second place with 25.14dB PSNR on the custom dataset, showing competitive results.

Abstract

目的镜面反射会干扰视觉感知系统,降低目标检测和三维重建等任务的性能。尽管现有高光去除方法已有进展,但大多依赖合成数据训练,难以有效处理真实场景中的反射。针对这一问题,本文构建了多光照条件下的汽车高光数据集,并提出掩码引导的高光去除网络(Mask Guided Transformer Unet,MG-TransUnet),显著提升高光去除效果。方法首先,本文构建了汽车高光数据集,覆盖多种光照条件及不同强度、分布的高光图像对,并提供高光掩码。并且提出协同优化的两阶段网络架构,第一阶段采用轻量级U形网络(U-shaped network,U-Net)实现高光检测,生成高光区域掩码;第二阶段将掩码作为空间先验,指导基于改进型窗口移位注意力(Shifted Window Attention Transformer,SWAT)的U-Net高光去除网络。通过端到端联合训练,实现检测与去除网络的参数协同优化。结果本文在自建数据集和三个公开数据集(PSD、SHIQ、SSHR)上进行了系统评估,与8种最先进方法进行对比。在PSD数据集上,本方法以26. 37dB的峰值信噪比(pe-ak signal-to-noise ratio,PSNR)位列第二;在SHIQ数据集上以35. 24dB的PSNR值优于第二名(34. 13dB)1. 11dB;在SSHR数据集上以36. 54dB的PSNR值领先第二名(36. 16dB)0. 38dB;在自建数据集上以25. 14dB的PSNR值优于第二名(25. 03dB)0. 11dB。此外,消融实验表明掩码引导网络和SWAT模块均能提升高光去除性能。结论本文构建了面向真实场景的汽车高光数据集,提出了基于窗口移位注意力机制的掩码引导网络(MG-TransUNet)。实验证明,本文方法在自建数据集上性能达到最优,并在多个公开数据集上取得了具有竞争力的表现。代码地址:https://github. com/chenWULUQI/MG-TransUnet

Ask AI
Helpful
Bookmark
Share
View Full Paper

Cite This Study

Jiacheng et al. (2026) studied this question.

synapsesocial.com/papers/69b3ab0002a1e69014ccba19https://doi.org/10.11834/jig.250353
Ask AI
Helpful
Bookmark
Share
View Full Paper