hellloxiaotian / Brdnet

Image denoising using deep CNN with batch renormalization（BRDNet）by Chunwei Tian, Yong Xu and Wangmeng Zuo is publised in Neural Networks, 2020. (https://www.sciencedirect.com/science/article/pii/S0893608019302394) and it is implemented by Keras.

This paper is pushed on home page of the Nueral Networks and BRDNet is collected by iHub，which is Chinese largest artificial intelligence platform. Also, it becomes a ESI highly cited paper. Additionally, it is reported by wechat public accounts at https://mp.weixin.qq.com/s/Jk6PlRBYorLI5FSa5xxOkw and https://mp.weixin.qq.com/s/dSCRx-6QW9bFDYQkDBGdLw.

This paper is the first paper via using enlaring the network width for addressing image denoising. Also, it is the first paper via deep network to address real noisy images of CC.

Absract

Deep convolutional neural networks (CNNs) have attracted great attention in the field of image denoising. However, there are two drawbacks: (1) It is very difficult to train a deeper CNN for denoising tasks, and (2) most of deeper CNNs suffer from performance saturation. In this paper, we report the design of a novel network called a batch-renormalization denoising network (BRDNet). Specifically, we combine two networks to increase the width of the network, and thus obtain more features. Because batch renormalization is fused into BRDNet, we can address the internal covariate shift and small mini-batchproblems. Residual learning is also adopted in a holistic way to facilitate network training. Dilated convolutions are exploited to extract more information for denoising tasks. Extensive experimental results show that BRDNet outperforms state-of-the-art image denoising methods. The code of BRDNet is accessible at http://www.yongxu.org/lunwen.html.

Requirements (Keras)

Tensorflow 1.3.0

Keras 2.0

Numpy

Opencv 2

Python 2.7

Cuda 8.0

Cudnn 7.5

Ubuntu 14.04

Commands

Training for gray noisy images

python mainimprovement.py

Training for color noisy images

python mainimprovement.py

Test for gray noisy images---test gray noisy image with noise level of 25

python mainimprovement.py --only_test True --pretrain 25/model_50.h5

Test for color noisy images---test color noisy image with noise level of 25

python mainimprovementcolor.py --only_test True --pretrain 25/model_50.h5

Training datasets

The training dataset of the gray noisy images is downloaded at https://pan.baidu.com/s/13jDZfayiM-KxLAOVWvFlUA

The training dataset of the color noisy images is downloaded at https://pan.baidu.com/s/1cx3ymsWLIT-YIiJRBza24Q

Network architecture

Resluts

Gaussian gray noisy image denoising

Average PSNR (dB) results of different methods on BSD68 dataset with noise levels of 15, 25 and 50.

PSNR (dB) results for different methods on 12 widely used images with noise levels of 15, 25 and 50.

Visual results for gray noisy images

Denoising results of one image from the BSD68 dataset with noise level 25 using for different methods: (a) original image, (b) noisy image /20.30 dB, (c) WNNM/29.75 dB, (d) E-PLL/29.59 dB, (e) TNRD/29.76 dB, (f) DnCNN/30.16 dB, (g) BM3D/29.53 dB, (h) IRCNN/30.07 dB, and(i) BRDNet/30.27 dB.

Denoising results of image “monar” from Set12 with noise level 50 using different methods: (a) original image, (b) noisy image/14.71 dB, (c) WNNM/26.32 dB, (d) EPLL/25.94dB, (e) TNRD/26.31 dB, (f) DnCNN/26.78 dB, (g) BM3D/25.82 dB, (h) IRCNN/26.61 dB, and(i) BRDNet/26.97 dB.

Gaussian color noisy image Denoising

Average PSNR (dB) results of different methods on the CBSD68, Kodak24, and McMaster datasets with noise levels of 15, 25, 35, 50, and 75.

Visual results for color noisy images

Denoising results for one color image from the McMaster dataset with noise level 35: (a) original image/ σ = 35, (b) noisy image/18.62 dB, (c) CBM3D/31.04 dB, (d) FFDNet/31.94dB, and (e) BRDNet/32.25 dB.

Denoising results for one color image from the Kodak24 dataset with noise level 60:(a) original image/ σ = 60, (b) noisy image/13.45 dB, (c) CBM3D/31.00 dB, (d) FFDNet/31.49 dB, and (e) BRDNet/31.85 dB.

Real noisy image denoising

PSNR (dB) results for different methods on real noisy images.

Complexity and complexity of different methods for image denoising

Complexity analysis of BRDNet, DnCNN and two DnCNNs.

Running time of different methods on an image different size

Running time for different methods in denoising images of sizes 256 × 256, 512 × 512, and 1024 × 1024.

If you want to cite this paper, please refer to the following format

1. Tian C, Xu Y, Zuo W. Image denoising using deep CNN with batch renormalization[J]. Neural Networks, 2020, 121: 461-473.

2. @article{tian2020image,

title={Image denoising using deep CNN with batch renormalization},

author={Tian, Chunwei and Xu, Yong and Zuo, Wangmeng},

journal={Neural Networks},

volume={121},

pages={461--473},

year={2020},

publisher={Elsevier}

}