Master of Science (M.S.), University of Dayton, 2021, Electrical Engineering

Noise is ubiquitous to practically all types of digital imaging systems. Low light color imaging is particularly challenging as the performance of demosaicking is affected by the presence of noise. Decoupling demosaicking and denoising tasks therefore results in artifacts. In this thesis, we address the low light color imaging problem by designing demosaicking in conjunction with denoising. Representing the RGB image as a combination of luminance and chrominance components, we derive a novel Bayer CFA joint demosaicking and denoising technique, based on a combination of wavelet-based demosaicking and Aitchison geometry modeling of wavelet-logarithm. The proposed demosaicking method is a minimum mean squared error estimate of the latent luminance and chrominance Aitchison variables, whose prior distribution is modeled as Gaussian scale mixtures. The resultant joint demosaicking-denoising method yields RGB image from noisy CFA data with image contrast details preserved while attenuating the noise. We verify the effectiveness of the proposed algorithm on a new 42 megapixel raw RGB sensor data.

Committee: Keigo Hirakawa (Advisor) Subjects: Electrical Engineering