Foveated Non-Local Means Denoising of Color Images, with Cross-Channel Paradigm.

Abstract

Foveation, a peculiarity of the HVS, is characterized by a sharp image having maximal acuity at the central part of the retina, the fovea. The acuity rapidly decreases towards the periphery of the visual field. Foveated imaging was recently investigated for the purpose of image denoising in the Foveated Non-local Means (FNLM) algorithm, and it was shown that for natural images the foveated self-similarity is a far more effective regularization prior than the conventional windowed self-similarity. Color images exhibit spectral redundancy across the R, G and B channels which can be exploited to reduce the effects of noise.

We extend the FNLM algorithm to the removal of additive white Gaussian noise from color images. The proposed Color-mixed Foveated NL-means algorithm, denominated as C-FNLM, implements the concept of foveated self-similarity, along with a cross-channel paradigm to exploit the correlation between color channels. The patch similarity is measured through an updated foveated distance for color images. In C-FNLM, we derive the explicit construction of an unified operator which explores the spatially variant nature of color perception in the HVS.

We develop a framework for designing the linear operator that simultaneously performs foveation and color mixing. Within this framework, we construct several parametrized families of the color-mixing operation. Our analysis shows that the color-mixed foveation is a far more effective regularity assumption than the windowing conventionally used in NL-means, especially for color image denoising where substantial improvement was observed in terms of contrast and sharpness. Moreover, the unified operator is introduced at a negligible cost in terms of the computational complexity.