4 个月前

多层小波-CNN图像复原

多层小波-CNN图像复原

摘要

在低层次视觉中,感受野大小与计算效率之间的权衡是一个关键问题。传统的卷积神经网络(CNNs)通常通过增加计算成本来扩大感受野。最近,空洞滤波被采用以解决这一问题。然而,它存在网格效应,导致生成的感受野仅是对输入图像的稀疏采样,呈现出棋盘状图案。本文提出了一种新颖的多级小波CNN(MWCNN)模型,旨在更好地平衡感受野大小与计算效率。通过改进的U-Net架构,在收缩子网络中引入了小波变换以减小特征图的尺寸。此外,还使用了一个额外的卷积层来减少特征图的通道数。在扩展子网络中,则部署了逆小波变换以重建高分辨率特征图。我们的MWCNN可以被视为空洞滤波和下采样的泛化形式,并可应用于多种图像恢复任务。实验结果清楚地表明了MWCNN在图像去噪、单幅图像超分辨率以及JPEG图像伪影去除方面的有效性。

基准测试

基准方法指标
grayscale-image-denoising-on-bsd68-sigma15MWCNN
PSNR: 31.86
grayscale-image-denoising-on-bsd68-sigma25MWCNN
PSNR: 29.41
grayscale-image-denoising-on-bsd68-sigma50MWCNN
PSNR: 26.53
grayscale-image-denoising-on-set12-sigma15MWCNN
PSNR: 33.15
grayscale-image-denoising-on-set12-sigma25MWCNN
PSNR: 30.79
grayscale-image-denoising-on-set12-sigma50MWCNN
PSNR: 27.74
grayscale-image-denoising-on-urban100-sigma15MWCNN
PSNR: 33.17
grayscale-image-denoising-on-urban100-sigma25MWCNN
PSNR: 30.66
grayscale-image-denoising-on-urban100-sigma50MWCNN
PSNR: 27.42
image-super-resolution-on-bsd100-2x-upscalingMWCNN
PSNR: 32.23
image-super-resolution-on-bsd100-3x-upscalingMWCNN
PSNR: 29.12
image-super-resolution-on-bsd100-4x-upscalingMWCNN
PSNR: 27.62
SSIM: 0.7355
image-super-resolution-on-set14-2x-upscalingMWCNN
PSNR: 33.7
image-super-resolution-on-set14-3x-upscalingMWCNN
PSNR: 30.16
image-super-resolution-on-set14-4x-upscalingMWCNN
PSNR: 28.41
SSIM: 0.7816
image-super-resolution-on-set5-2x-upscalingMWCNN
PSNR: 37.91
image-super-resolution-on-set5-3x-upscalingMWCNN
PSNR: 34.17
image-super-resolution-on-urban100-2xMWCNN
PSNR: 32.3
image-super-resolution-on-urban100-3xMWCNN
PSNR: 28.13
image-super-resolution-on-urban100-4xMWCNN
PSNR: 26.27
SSIM: 0.7890
jpeg-artifact-correction-on-classic5-qualityMWCNN
PSNR: 30.01
jpeg-artifact-correction-on-classic5-quality-1MWCNN
PSNR: 32.16
jpeg-artifact-correction-on-classic5-quality-2MWCNN
PSNR: 33.43
jpeg-artifact-correction-on-classic5-quality-3MWCNN
PSNR: 34.27
jpeg-artifact-correction-on-icb-quality-10MWCNN
PSNR: 30.76
PSNR-B: 31.21
SSIM: 0.779
jpeg-artifact-correction-on-icb-quality-10-1MWCNN
PSNR: 34.12
PSNR-B: 34.06
SSIM: 0.884
jpeg-artifact-correction-on-icb-quality-20MWCNN
PSNR: 32.79
PSNR-B: 33.32
SSIM: 0.812
jpeg-artifact-correction-on-icb-quality-20-1MWCNN
PSNR: 36.56
PSNR-B: 36.44
SSIM: 0.902
jpeg-artifact-correction-on-icb-quality-30MWCNN
PSNR: 34.11
PSNR-B: 34.69
SSIM: 0.845
jpeg-artifact-correction-on-live1-quality-10MWCNN
PSNR: 27.45
PSNR-B: 27.44
SSIM: 0.808
jpeg-artifact-correction-on-live1-quality-10-1MWCNN
PSNR: 29.69
PSNR-B: 29.39
SSIM: 0.8357
jpeg-artifact-correction-on-live1-quality-20MWCNN
PSNR: 29.80
PSNR-B: 29.78
SSIM: 0.877
jpeg-artifact-correction-on-live1-quality-20-1MWCNN
PSNR: 32.04
PSNR-B: 31.83
SSIM: 0.8989
jpeg-artifact-correction-on-live1-quality-30-1MWCNN
PSNR: 33.45
jpeg-artifact-correction-on-live1-quality-40MWCNN
PSNR: 34.45

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
多层小波-CNN图像复原 | 论文 | HyperAI超神经