Beyond a Gaussian Denoiser: Residual Learning of Deep CNN for Image Denoising

13 Aug 2016  ·  Kai Zhang, WangMeng Zuo, Yunjin Chen, Deyu Meng, Lei Zhang ·

Discriminative model learning for image denoising has been recently attracting considerable attentions due to its favorable denoising performance. In this paper, we take one step forward by investigating the construction of feed-forward denoising convolutional neural networks (DnCNNs) to embrace the progress in very deep architecture, learning algorithm, and regularization method into image denoising... Specifically, residual learning and batch normalization are utilized to speed up the training process as well as boost the denoising performance. Different from the existing discriminative denoising models which usually train a specific model for additive white Gaussian noise (AWGN) at a certain noise level, our DnCNN model is able to handle Gaussian denoising with unknown noise level (i.e., blind Gaussian denoising). With the residual learning strategy, DnCNN implicitly removes the latent clean image in the hidden layers. This property motivates us to train a single DnCNN model to tackle with several general image denoising tasks such as Gaussian denoising, single image super-resolution and JPEG image deblocking. Our extensive experiments demonstrate that our DnCNN model can not only exhibit high effectiveness in several general image denoising tasks, but also be efficiently implemented by benefiting from GPU computing. read more

PDF Abstract

Datasets


Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Image Super-Resolution BSD100 - 2x upscaling DnCNN-3 PSNR 31.9 # 12
Image Super-Resolution BSD100 - 3x upscaling DnCNN-3 PSNR 28.85 # 9
Image Super-Resolution BSD100 - 4x upscaling DnCNN-3 PSNR 27.29 # 35
SSIM 0.7253 # 36
Color Image Denoising BSD68 sigma15 DnCNN-3 PSNR 31.46 # 4
Grayscale Image Denoising BSD68 sigma25 DnCNN PSNR 29.23 # 7
Color Image Denoising BSD68 sigma25 DnCNN-3 PSNR 29.02 # 4
Color Image Denoising CBSD68 sigma35 DnCNN-B* PSNR 28.74 # 4
JPEG Artifact Correction Classic5 (Quality 10 Grayscale) DnCNN-3 PSNR 29.4 # 5
JPEG Artifact Correction Classic5 (Quality 20 Grayscale) DnCNN-3 PSNR 31.63 # 4
JPEG Artifact Correction Classic5 (Quality 30 Grayscale) DnCNN-3 PSNR 32.91 # 4
JPEG Artifact Correction Classic5 (Quality 40 Grayscale) DnCNN-3 PSNR 33.77 # 3
JPEG Artifact Correction Live1 (Quality 10 Grayscale) DnCNN-3 PSNR 29.19 # 9
JPEG Artifact Correction LIVE1 (Quality 20 Grayscale) DnCNN-3 PSNR 31.59 # 9
JPEG Artifact Correction LIVE1 (Quality 30 Grayscale) DnCNN-3 PSNR 32.98 # 4
JPEG Artifact Correction LIVE1 (Quality 40 Grayscale) DnCNN-3 PSNR 33.96 # 3
Image Super-Resolution Set14 - 2x upscaling DnCNN-3 PSNR 33.03 # 15
Image Super-Resolution Set14 - 3x upscaling DnCNN-3 PSNR 29.81 # 7
Image Super-Resolution Set14 - 4x upscaling DnCNN-3 PSNR 28.04 # 37
SSIM 0.7672 # 40
Image Super-Resolution Set5 - 2x upscaling DnCNN-3 PSNR 37.58 # 17
Image Super-Resolution Set5 - 3x upscaling DnCNN-3 PSNR 33.75 # 12
Image Super-Resolution Set5 - 4x upscaling DnCNN-3 PSNR 31.4 # 40
SSIM 0.8845 # 39
Image Super-Resolution Urban100 - 2x upscaling DnCNN-3 PSNR 30.74 # 15
Image Super-Resolution Urban100 - 3x upscaling DnCNN-3 PSNR 27.15 # 7
Image Super-Resolution Urban100 - 4x upscaling DnCNN-3 PSNR 25.2 # 33
SSIM 0.7521 # 32
Grayscale Image Denoising Urban100 sigma15 DnCNN PSNR 32.67 # 4
Grayscale Image Denoising Urban100 sigma25 DnCNN PSNR 29.97 # 5

Results from Other Papers


Task Dataset Model Metric Name Metric Value Rank Source Paper Compare
Denoising Darmstadt Noise Dataset CDnCNN-B PSNR 32.43 # 8

Methods


No methods listed for this paper. Add relevant methods here