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.
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.
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
PSNR (dB) results for different methods on 12 widely used images with noise levels of 15, 25 and 50.
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.
Average PSNR (dB) results of different methods on the CBSD68, Kodak24, and McMaster datasets with noise levels of 15, 25, 35, 50, and 75.
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.
Running time for different methods in denoising images of sizes 256 × 256, 512 × 512, and 1024 × 1024.
