3 个月前

ML-CrAIST:基于多尺度高低频信息的跨黑注意力图像超分辨率Transformer

ML-CrAIST:基于多尺度高低频信息的跨黑注意力图像超分辨率Transformer

摘要

近年来,Transformer架构在单图像超分辨率任务中引起了广泛关注,并展现出显著的性能提升。然而,现有模型过度依赖网络提取图像高层语义细节的能力,却忽视了多尺度图像细节以及网络内部中间特征的有效利用。此外,研究发现,与低频区域相比,图像中的高频区域在超分辨率任务中具有更高的复杂性。为此,本文提出一种基于Transformer的超分辨率架构——ML-CrAIST,该方法通过融合多尺度的低频与高频信息来弥补这一不足。与以往大多数仅在空间维度或通道维度上操作的模型不同,本文同时引入空间自注意力与通道自注意力机制,能够协同建模像素在空间维度和通道维度上的交互关系,充分挖掘空间与通道轴之间的内在关联性。此外,我们设计了一种用于超分辨率任务的交叉注意力模块,用于探索低频与高频信息之间的相关性。定量与定性实验结果表明,所提出的ML-CrAIST在多个基准数据集上显著优于当前最先进的超分辨率方法(例如,在Manga109 ×4任务上提升达0.15 dB)。代码已开源,访问地址:https://github.com/Alik033/ML-CrAIST。

代码仓库

alik033/ml-craist
官方
pytorch
GitHub 中提及

基准测试

基准方法指标
image-super-resolution-on-2x-upscalingML-CrAIST
#params (K): 1259
FLOPs(G): 165.7
image-super-resolution-on-2x-upscalingML-CrAIST-Li
#params (K): 743
FLOPs(G): 97.2
image-super-resolution-on-3x-upscalingML-CrAIST
#params (K): 1268
FLOPs(G): 84.1
image-super-resolution-on-3x-upscalingML-CrAIST-Li
#params (K): 749
FLOPs(G): 49.6
image-super-resolution-on-4x-upscalingML-CrAIST-Li
#params (K): 758
FLOPs(G): 25.5
image-super-resolution-on-4x-upscalingML-CrAIST
#params (K): 1280
FLOPs(G): 42.9
image-super-resolution-on-b100-2x-upscalingML-CrAIST
SSIM: 0.9022
image-super-resolution-on-b100-2x-upscalingML-CrAIST-Li
PSNR: 32.36
SSIM: 0.902
image-super-resolution-on-b100-3x-upscalingML-CrAIST
SSIM: 0.8111
image-super-resolution-on-b100-3x-upscalingML-CrAIST-Li
PSNR: 29.28
SSIM: 0.8106
image-super-resolution-on-b100-4x-upscalingML-CrAIST
PSNR: 27.78
SSIM: 0.7446
image-super-resolution-on-b100-4x-upscalingML-CrAIST-Li
PSNR: 27.73
image-super-resolution-on-manga109-2xML-CrAIST
PSNR: 39.26
SSIM: 0.9786
image-super-resolution-on-manga109-2xML-CrAIST-Li
PSNR: 39.23
SSIM: 0.9785
image-super-resolution-on-manga109-3xML-CrAIST-Li
PSNR: 34.26
SSIM: 0.9492
image-super-resolution-on-manga109-3xML-CrAIST
PSNR: 34.42
SSIM: 0.9501
image-super-resolution-on-manga109-4xML-CrAIST
PSNR: 31.17
SSIM: 0.9176
image-super-resolution-on-manga109-4xML-CrAIST-Li
PSNR: 31.11
SSIM: 0.9162
image-super-resolution-on-set14-2x-upscalingML-CrAIST-Li
PSNR: 33.64
SSIM: 0.9213
image-super-resolution-on-set14-2x-upscalingML-CrAIST
PSNR: 33.77
SSIM: 0.922
image-super-resolution-on-set14-3x-upscalingML-CrAIST
PSNR: 30.39
SSIM: 0.8488
image-super-resolution-on-set14-3x-upscalingML-CrAIST-Li
PSNR: 30.23
SSIM: 0.8474
image-super-resolution-on-set14-4x-upscalingML-CrAIST-Li
PSNR: 28.4
SSIM: 0.7863
image-super-resolution-on-set14-4x-upscalingML-CrAIST
PSNR: 28.53
SSIM: 0.7895
image-super-resolution-on-set5-2x-upscalingML-CrAIST
PSNR: 38.19
SSIM: 0.9617
image-super-resolution-on-set5-2x-upscalingML-CrAIST-Li
PSNR: 38.15
SSIM: 0.9615
image-super-resolution-on-set5-3x-upscalingML-CrAIST
PSNR: 34.7
SSIM: 0.9302
image-super-resolution-on-set5-3x-upscalingML-CrAIST-Li
PSNR: 34.58
SSIM: 0.9294
image-super-resolution-on-set5-4x-upscalingML-CrAIST
PSNR: 32.36
SSIM: 0.8984
image-super-resolution-on-set5-4x-upscalingML-CrAIST-Li
PSNR: 32.15
SSIM: 0.8962
image-super-resolution-on-urban100-2xML-CrAIST-Li
PSNR: 32.93
SSIM: 0.9361
image-super-resolution-on-urban100-2xML-CrAIST
PSNR: 33.04
SSIM: 0.937
image-super-resolution-on-urban100-3xML-CrAIST-Li
PSNR: 28.73
SSIM: 0.8651
image-super-resolution-on-urban100-3xML-CrAIST
PSNR: 28.89
SSIM: 0.8676
image-super-resolution-on-urban100-4xML-CrAIST
PSNR: 26.68
SSIM: 0.8057
image-super-resolution-on-urban100-4xML-CrAIST-Li
PSNR: 26.53
SSIM: 0.8019

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
ML-CrAIST:基于多尺度高低频信息的跨黑注意力图像超分辨率Transformer | 论文 | HyperAI超神经