Efficient cascading of multi-domain image Gaussian noise filters
- PDF / 4,470,202 Bytes
- 13 Pages / 595.276 x 790.866 pts Page_size
- 101 Downloads / 237 Views
ORIGINAL RESEARCH PAPER
Efficient cascading of multi‑domain image Gaussian noise filters Meisam Rakhshanfar1 · Maria A. Amer1 Received: 28 September 2018 / Accepted: 9 March 2019 © Springer-Verlag GmbH Germany, part of Springer Nature 2019
Abstract Image denoising is a well explored but still an active research topic. The focus is usually on achieving higher numerical quality which is theoretically interesting, however, often the factor of computation cost is not considered. Our idea is to employ different image Gaussian noise filters to construct an effective image denoiser, where the deficiency of each filter is compensated with others, while a wide variation of quality versus speed can be achieved. We integrate filters using different cascaded forms and show that if two filters use uncorrelated features, their cascaded form provides a higher quality than each separately. We start with easy-to-implement filters employing pixel- and frequency-domain with different kernel size to construct a fast yet high-quality multi-domain denoiser. Then, we propose more complex denoisers by integrating our cascaded multi-domain denoiser to other state-of-the-art denoising methods. Simulations show that the quality of proposed multi-domain denoiser is significantly higher than its building-blocks. We also show that the proposed multi-domain denoiser can be integrated to state-of-the-art denoisers to from a more effective denoiser, while adding negligible complexity. Keywords Image denoising · Multi-domain denoiser · Cascaded filters · White noise
1 Introduction Image denoising plays an important role in improving visual perception of images and in many other image applications. The theory behind image denoising is based on analyzing neighboring pixels inside a certain window to estimate the true value of the pixels. As the size of cameras and their lenses decreases (e.g., mobile devices) noise becomes more intensive. Improvement of an image degraded by noise has been studied for decades and new approaches pushing this field’s boundary by achieving higher qualities but with significant computational cost. Even with many advances in digital hardware, most of the state-of-the-art are impractical in consumer electronic applications, even with high-end processors. Additionally, the amount of information that can be extracted inside a certain window is restricted and, thus, the achievable numerical quality in spatial filtering is limited [1]. These numerical improvements are interesting, however, they often cannot be detected by the human eye [2]. Image denoising benefits from high correlation between neighboring pixels. Since such correlation at edges is low, * Maria A. Amer 1
Department of Electrical and Computer Engineering, Concordia University, Montreal, QC, Canada
many edge-stopping pixel-domain filters, such as bilateral filtering and anisotropic diffusion, have been introduced. Ideal denoiser suppresses the noise while preserves textures and edges. Image textures are often repeated and thus many approaches [such as b
Data Loading...
