Acceleration of Non-Linear Image Filters, and Multi-Frame Image Denoising
Date of Award
Ph.D. in Engineering
Department of Electrical and Computer Engineering
Advisor: Keigo Hirakawa
This dissertation is comprised of four novel contributions. First, we propose new implementations of Monte-Carlo-based bilateral filter and non-local means whose per-pixel complexity is approximately invariant to the color dimension, window size, and block size. We reduce complexity by combining the random filtering of multiple color channels that approximate the non-linear behavior of the bilateral filter into a single convolution operation. We extend this work to a non-linear filter called Non-Local Means. In the second part, we propose "convolutional distance transform"-- efficient implementations of distance transform. Specifically, we leverage approximate minimum functions to rewrite the distance transform in terms of convolution operators, reducing the complexity to N logN. Third, we propose a novel method for multi-frame image denoising in mobile phones. We developed a method to register noisy image frames by estimating the camera motion using both image and inertial measurements. Lastly, we develop a new framework for multi-frame image denoising using noisy image statistics of one frame to design an optimal denoising filter for the second frame. The algorithm is provably optimal in minimum mean squared error estimation sense as well as in wavelet structural similarity metric sense.
Electrical Engineering, bilateral filter, non-local means, noise2noise, multi-frame denoising
Copyright 2019, author
Karam, Christina Maria, "Acceleration of Non-Linear Image Filters, and Multi-Frame Image Denoising" (2019). Graduate Theses and Dissertations. 6673.