4 个月前

SwinFIR:基于快速傅里叶卷积和改进训练的图像超分辨率SwinIR重访

SwinFIR:基于快速傅里叶卷积和改进训练的图像超分辨率SwinIR重访

摘要

基于Transformer的方法由于其在建模长距离依赖方面相比卷积神经网络(CNN)方法具有更强的能力,因此在图像恢复领域取得了令人印象深刻的性能。然而,诸如SwinIR等方法采用了基于窗口的局部注意力策略来平衡性能和计算开销,这限制了早期层中使用大感受野来捕获全局信息和建立长距离依赖。为了进一步提高捕捉全局信息的效率,在本研究中,我们提出了SwinFIR,通过替换具有全图感受野的快速傅里叶卷积(FFC)组件来扩展SwinIR。我们还重新审视了其他先进的技术,如数据增强、预训练和特征融合,以改进图像重建的效果。我们的特征融合方法能够在不增加训练和测试时间的情况下显著提升模型的性能。我们在多个流行的大型基准数据集上应用了该算法,并与现有方法相比达到了最先进的性能。例如,我们的SwinFIR在Manga109数据集上实现了32.83分贝的峰值信噪比(PSNR),比最先进的SwinIR方法高出0.8分贝。

代码仓库

IMPLabUniPr/swin2-mose
pytorch
GitHub 中提及

基准测试

基准方法指标
image-super-resolution-on-bsd100-2x-upscalingSwinFIR
PSNR: 32.64
SSIM: 0.9054
image-super-resolution-on-bsd100-2x-upscalingHAT_FIR
PSNR: 32.71
image-super-resolution-on-bsd100-3x-upscalingHAT_FIR
PSNR: 29.6
image-super-resolution-on-bsd100-3x-upscalingSwinFIR
PSNR: 29.55
SSIM: 0.8169
image-super-resolution-on-bsd100-4x-upscalingHAT_FIR
PSNR: 28.07
image-super-resolution-on-bsd100-4x-upscalingSwinFIR
PSNR: 28.03
SSIM: 0.7520
image-super-resolution-on-manga109-2xHAT_FIR
PSNR: 40.77
image-super-resolution-on-manga109-2xSwinFIR
PSNR: 40.61
SSIM: 0.9816
image-super-resolution-on-manga109-3xHAT_FIR
PSNR: 35.92
image-super-resolution-on-manga109-3xSwinFIR
PSNR: 35.77
SSIM: 0.9563
image-super-resolution-on-manga109-4xHAT_FIR
PSNR: 33.03
image-super-resolution-on-manga109-4xSwinFIR
PSNR: 32.83
SSIM: 0.9314
image-super-resolution-on-set14-2x-upscalingHAT_FIR
PSNR: 35.17
image-super-resolution-on-set14-2x-upscalingSwinFIR
PSNR: 34.93
SSIM: 0.9276
image-super-resolution-on-set14-3x-upscalingSwinFIR
PSNR: 31.24
SSIM: 0.8566
image-super-resolution-on-set14-3x-upscalingHAT_FIR
PSNR: 31.37
image-super-resolution-on-set14-4x-upscalingHAT_FIR
PSNR: 29.44
image-super-resolution-on-set14-4x-upscalingSwinFIR
PSNR: 29.36
SSIM: 0.7993
image-super-resolution-on-set5-2x-upscalingSwinFIR
PSNR: 38.65
SSIM: 0.9633
image-super-resolution-on-set5-2x-upscalingHAT_FIR
PSNR: 38.74
image-super-resolution-on-set5-3x-upscalingHAT_FIR
PSNR: 35.21
image-super-resolution-on-set5-3x-upscalingSwinFIR
PSNR: 35.15
SSIM: 0.9330
image-super-resolution-on-urban100-2xSwinFIR
PSNR: 34.57
SSIM: 0.9473
image-super-resolution-on-urban100-2xHAT_FIR
PSNR: 34.94
image-super-resolution-on-urban100-3xSwinFIR
PSNR: 30.43
SSIM: 0.8913
image-super-resolution-on-urban100-3xHAT_FIR
PSNR: 30.77
image-super-resolution-on-urban100-4xHAT_FIR
PSNR: 28.43
image-super-resolution-on-urban100-4xSwinFIR
PSNR: 28.12
SSIM: 0.8393
stereo-image-super-resolution-on-flickr1024-1SwinFIRSSR
PSNR: 30.14
stereo-image-super-resolution-on-flickr1024-2SwinFIRSSR
PSNR: 24.29
stereo-image-super-resolution-on-kitti2012-2x-1SwinFIRSSR
PSNR: 31.79
stereo-image-super-resolution-on-kitti2012-2x-2SwinFIRSSR
PSNR: 31.79
stereo-image-super-resolution-on-kitti2012-4xSwinFIRSSR
PSNR: 27.16
stereo-image-super-resolution-on-kitti2015-4xSwinFIRSSR
PSNR: 26.89
stereo-image-super-resolution-on-middleburySwinFIRSSR
PSNR: 30.44

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
SwinFIR:基于快速傅里叶卷积和改进训练的图像超分辨率SwinIR重访 | 论文 | HyperAI超神经