Loading [a11y]/accessibility-menu.js
Revisiting DCT in Deep Learning Era: An Initial Denoising Application | IEEE Conference Publication | IEEE Xplore

Revisiting DCT in Deep Learning Era: An Initial Denoising Application


Abstract:

In the classical era of image denoising, the methods working in transform domain have achieved high performance results. However, most of the deep neural networks that ha...Show More

Abstract:

In the classical era of image denoising, the methods working in transform domain have achieved high performance results. However, most of the deep neural networks that have been proposed in the last decade and have shown better noise removal performance try to denoise the noisy image in pixel domain. In deep learning literature, there are few deep neural networks that work in transform domain. Most of them have not chosen the discrete cosine transform (DCT), which is known to provide a very good representation for most images. This is the result of convolution layer, which is often used in deep networks, searching in vain for a relationship between neighboring values of an image’s uncorrelated global and block DCT coefficients. On the other hand, it is known that working with transform coefficients of overlapped image blocks improves noise removal performance. Recent studies have shown that the convolution of an image with 2D DCT basis images is a meaningful ordering of the DCT coefficients of overlapping image blocks. Hence, in this paper, a deep neural network is proposed to remove noise in the DCT domain. Experiments on color images indicate that the proposed network is quantitatively and qualitatively successful in noise removal.
Date of Conference: 15-18 May 2024
Date Added to IEEE Xplore: 23 July 2024
ISBN Information:
Print on Demand(PoD) ISSN: 2165-0608
Conference Location: Mersin, Turkiye