4 个月前

Retinexformer:基于Retinex的一阶段Transformer低光图像增强

Retinexformer:基于Retinex的一阶段Transformer低光图像增强

摘要

在增强低光图像时,许多深度学习算法都基于Retinex理论。然而,Retinex模型并未考虑暗部隐藏的噪声或由增亮过程引入的噪声。此外,这些方法通常需要繁琐的多阶段训练流程,并依赖于卷积神经网络,这在捕捉长距离依赖关系方面显示出局限性。本文提出了一种简单而有原则的一阶段Retinex基础框架(One-stage Retinex-based Framework,简称ORF)。ORF首先估计光照信息以照亮低光图像,然后恢复图像中的噪声以生成增强后的图像。我们设计了一种光照引导的Transformer(Illumination-Guided Transformer,简称IGT),该模型利用光照表示来指导不同光照条件下区域之间的非局部相互作用建模。通过将IGT集成到ORF中,我们得到了我们的算法——Retinexformer。全面的定量和定性实验表明,我们的Retinexformer在十三个基准测试中显著优于现有最先进方法。用户研究和低光物体检测应用也揭示了我们方法潜在的实际价值。代码、模型和结果可在https://github.com/caiyuanhao1998/Retinexformer 获取。

基准测试

基准方法指标
image-enhancement-on-mit-adobe-5kRetinexformer
PSNR on proRGB: 25.98
PSNR on sRGB: 24.94
SSIM on proRGB: 0.957
SSIM on sRGB: 0.907
low-light-image-deblurring-and-enhancement-onRetinexFormer
Average PSNR: 22.904
LPIPS: 0.236
SSIM: 0.824
low-light-image-enhancement-on-dicmRetinexformer
User Study Score: 3.71
low-light-image-enhancement-on-limeRextinexformer
User Study Score: 4.3
low-light-image-enhancement-on-lolRetinexformer_
Average PSNR: 27.18
FLOPS (G): 15.57
Params (M): 1.61
SSIM: 0.850
low-light-image-enhancement-on-lolRetinexformer
Average PSNR: 25.16
FLOPS (G): 15.57
Params (M): 1.61
SSIM: 0.845
low-light-image-enhancement-on-lol-v2Retinexformer
Average PSNR: 22.80
SSIM: 0.840
low-light-image-enhancement-on-lol-v2-1Retinexformer
PSNR: 25.67
SSIM: 0.939
low-light-image-enhancement-on-lolv2Retinexformer
Average PSNR: 27.71
SSIM: 0.856
low-light-image-enhancement-on-lolv2-1Retinexformer
Average PSNR: 29.04
SSIM: 0.939
low-light-image-enhancement-on-mefRetinexformer
User Study Score: 3.91
low-light-image-enhancement-on-mit-adobe-1Retinexformer
PSNR: 24.94
SSIM: 0.907
low-light-image-enhancement-on-npeRetinexformer
User Study Score: 4.17
low-light-image-enhancement-on-sdsd-indoorRetinexformer
PSNR: 29.77
low-light-image-enhancement-on-sdsd-outdoorRetinexformer
PSNR: 29.84
low-light-image-enhancement-on-sidRetinexformer
PSNR: 24.44
SSIM: 0.680
low-light-image-enhancement-on-smidRetinexformer
PSNR: 29.15
low-light-image-enhancement-on-vvRextinexformer
User Study Score: 3.61
photo-retouching-on-mit-adobe-5kRetinexformer
PSNR: 24.94
SSIM: 0.907

用 AI 构建 AI

从想法到上线——通过免费 AI 协同编程、开箱即用的环境和市场最优价格的 GPU 加速您的 AI 开发

AI 协同编程
即用型 GPU
最优价格
立即开始

Hyper Newsletters

订阅我们的最新资讯
我们会在北京时间 每周一的上午九点 向您的邮箱投递本周内的最新更新
邮件发送服务由 MailChimp 提供
Retinexformer:基于Retinex的一阶段Transformer低光图像增强 | 论文 | HyperAI超神经