TUTCRIS

TY - JOUR

T1 - DCT-based denoising in multichannel imaging with reference

AU - Lukin, V.

AU - Abramov, S.

AU - Abramova, V.

AU - Astola, J.

AU - Egiazarian, K.

N1 - EXT="Lukin, V."

PY - 2016

Y1 - 2016

N2 - A task of denoising of a component image of multichannel data is considered in this paper assuming that a reference (noise-free) image is available. We propose a denoising approach based on three-dimensional (3D) discrete cosine transform (DCT) applied in blocks. We show that a use of a reference image allows improving the denoising performance (measured by different quality metrics) although it depends on several factors such as a choice of the reference and the way it is pre-processed. One of the most important requirements to achieve a good performance is a similarity between to be processed and the reference images. A high cross-correlation between them is a necessary but not sufficient condition. These images should have also close dynamic range. If all these requirements are satisfied by an appropriate choice or by preprocessing of the reference, improvements of the metrics PSNR and PSNR-HVS-M can be up to 3...5 dB compared to the component-wise DCT-based image denoising. We also analyze and process real-life hyperspectral images and provide examples showing efficiency of filtering noisy component images using other components with high signal-to-noise ratios as references.

AB - A task of denoising of a component image of multichannel data is considered in this paper assuming that a reference (noise-free) image is available. We propose a denoising approach based on three-dimensional (3D) discrete cosine transform (DCT) applied in blocks. We show that a use of a reference image allows improving the denoising performance (measured by different quality metrics) although it depends on several factors such as a choice of the reference and the way it is pre-processed. One of the most important requirements to achieve a good performance is a similarity between to be processed and the reference images. A high cross-correlation between them is a necessary but not sufficient condition. These images should have also close dynamic range. If all these requirements are satisfied by an appropriate choice or by preprocessing of the reference, improvements of the metrics PSNR and PSNR-HVS-M can be up to 3...5 dB compared to the component-wise DCT-based image denoising. We also analyze and process real-life hyperspectral images and provide examples showing efficiency of filtering noisy component images using other components with high signal-to-noise ratios as references.

KW - 3D-DCT

KW - Denoising

KW - Image processing

KW - Multichannel images

KW - Quality metrics

KW - Reference image

UR - http://www.scopus.com/inward/record.url?scp=84995426770&partnerID=8YFLogxK

M3 - Article

VL - 75

SP - 1167

EP - 1191

JO - Telecommunications and Radio Engineering

JF - Telecommunications and Radio Engineering

SN - 0040-2508

IS - 13

ER -