search for


Noise reduction using patch-based CNN in images
Journal of the Korean Data & Information Science Society 2019;30:349-63
Published online March 31, 2019;
© 2019 Korean Data and Information Science Society.

Kwanghae Heo1 · Dong Hoon Lim2

12Department of Information Statistics, Gyeongsang National University
Correspondence to: 2 Corresponding author: Professor and RINS, Department of Information Statistics, Gyeongsang National University, Jinju 52828, Korea. E-mail:
This research was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (Ministry of Science and ICT) (No. NRF-2017R1E1A1A03071057).
Received February 6, 2019; Revised March 4, 2019; Accepted March 10, 2019.
This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License ( which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.
Noise reduction problem in images still prevails as a challenge in the field of image processing such as pattern recognition, image compression, edge detection and image segmentation. Addressing this issue, this paper presents a novel deep learning approach based on a Convolutional Neural Network (CNN). CNN has shown excellent performance in computer vision problems such as image recognition, object recognition, and face recognition, but little has been discussed in light of the importance of noise reduction in images. Until now, noise reduction in the images has been used with filters designed under the assumption that it has specific distribution characteristics. In this case, the use of filters that do not satisfy the assumption leads to significant performance degradation. In this paper, CNN is applied on patches of images in a way that is available without prior information about noise. The restored image is obtained by the weighted average of the corresponding pixels in overlapping patches. In CNN, parameter optimization is done by the Adam algorithm that is adaptable to noise data. We considered both Gaussian noise and impulse noise to test the performance of our CNN model. Experimental results on several images show that the patch-based CNN model yields significantly superior image quality and better MAE (mean absolute error) and PSNR (peak signal-to-noise ratio).
Keywords : Convolutional neural network (CNN), deep learning, noise reduction, noisy image.