3 个月前

在图像超分辨率Transformer中激活更多像素

在图像超分辨率Transformer中激活更多像素

摘要

基于Transformer的方法在低层视觉任务(如图像超分辨率)中展现出卓越的性能。然而,通过归因分析我们发现,现有网络在利用输入信息时,其空间感受野仍受到显著限制,这表明Transformer在现有架构中的潜力尚未得到充分挖掘。为激活更多输入像素以实现更优的重建效果,本文提出一种新型混合注意力Transformer(Hybrid Attention Transformer, HAT)。该方法融合了通道注意力与基于窗口的自注意力机制,充分发挥二者优势:前者能够有效利用全局统计信息,后者则具备强大的局部拟合能力。此外,为更有效地聚合跨窗口信息,我们引入了重叠交叉注意力模块,以增强相邻窗口特征之间的交互。在训练阶段,我们进一步采用同任务预训练策略,以进一步挖掘模型的潜力。大量实验验证了所提模块的有效性,同时通过模型规模扩展实验表明,该方法在任务性能上可实现显著提升。整体方法在各项指标上显著优于现有最先进方法,性能提升超过1 dB。相关代码与模型已开源,地址为:https://github.com/XPixelGroup/HAT。

代码仓库

chxy95/hat
官方
pytorch
GitHub 中提及
xpixelgroup/hat
官方
pytorch
GitHub 中提及

基准测试

基准方法指标
image-super-resolution-on-bsd100-2x-upscalingHAT-L
PSNR: 32.74
SSIM: 0.9066
image-super-resolution-on-bsd100-2x-upscalingHAT
PSNR: 32.69
SSIM: 0.9060
image-super-resolution-on-bsd100-3x-upscalingHAT
PSNR: 29.59
SSIM: 0.8177
image-super-resolution-on-bsd100-3x-upscalingHAT-L
PSNR: 29.63
SSIM: 0.8191
image-super-resolution-on-bsd100-4x-upscalingHAT
PSNR: 28.05
SSIM: 0.7534
image-super-resolution-on-bsd100-4x-upscalingHAT-L
PSNR: 28.09
SSIM: 0.7551
image-super-resolution-on-manga109-2xHAT-L
PSNR: 41.01
SSIM: 0.9831
image-super-resolution-on-manga109-2xHAT
PSNR: 40.71
SSIM: 0.9819
image-super-resolution-on-manga109-3xHAT
PSNR: 35.84
SSIM: 0.9567
image-super-resolution-on-manga109-3xHAT-L
PSNR: 36.02
SSIM: 0.9576
image-super-resolution-on-manga109-4xHAT-L
PSNR: 33.09
SSIM: 0.9335
image-super-resolution-on-manga109-4xHAT
PSNR: 32.87
SSIM: 0.9319
image-super-resolution-on-set14-2x-upscalingHAT-L
PSNR: 35.29
SSIM: 0.9293
image-super-resolution-on-set14-2x-upscalingHAT
PSNR: 35.13
SSIM: 0.9282
image-super-resolution-on-set14-3x-upscalingHAT
PSNR: 31.33
SSIM: 0.8576
image-super-resolution-on-set14-3x-upscalingHAT-L
PSNR: 31.47
SSIM: 0.8584
image-super-resolution-on-set14-4x-upscalingHAT
PSNR: 29.38
SSIM: 0.8001
image-super-resolution-on-set14-4x-upscalingHAT-L
PSNR: 29.47
SSIM: 0.8015
image-super-resolution-on-set5-2x-upscalingHAT
PSNR: 38.73
SSIM: 0.9637
image-super-resolution-on-set5-2x-upscalingHAT-L
PSNR: 38.91
SSIM: 0.9646
image-super-resolution-on-set5-3x-upscalingHAT
PSNR: 35.16
SSIM: 0.9335
image-super-resolution-on-set5-3x-upscalingHAT-L
PSNR: 35.28
SSIM: 0.9345
image-super-resolution-on-set5-4x-upscalingHAT-L
PSNR: 33.30
SSIM: 0.9083
image-super-resolution-on-urban100-2xHAT
PSNR: 34.81
SSIM: 0.9489
image-super-resolution-on-urban100-2xHAT-L
PSNR: 35.09
SSIM: 0.9505
image-super-resolution-on-urban100-3xHAT
PSNR: 30.70
SSIM: 0.8949
image-super-resolution-on-urban100-3xHAT-L
PSNR: 30.92
SSIM: 0.8981
image-super-resolution-on-urban100-4xHAT-L
PSNR: 28.60
SSIM: 0.8498
image-super-resolution-on-urban100-4xHAT
PSNR: 28.37
SSIM: 0.8447

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
在图像超分辨率Transformer中激活更多像素 | 论文 | HyperAI超神经