Date of Award

8-2018

Degree Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Electrical and Computer Engineering

Committee Chair

Stanley Chan

Committee Member 1

Charles A. Bouman

Committee Member 2

Mireille Boutin

Committee Member 3

Amy R. Reibman

Abstract

Feature learning is a technique to automatically extract features from raw data. It is widely used in areas such as computer vision, image processing, data mining and natural language processing. In this thesis, we are interested in the computational aspects of feature learning. We focus on rank matrix and tensor factorization and deep neural network models for image denoising.

With respect to matrix and tensor factorization, we first present a technique to speed up alternating least squares (ALS) and gradient descent (GD) − two commonly used strategies for tensor factorization. We introduce an efficient, scalable and distributed algorithm that addresses the data explosion problem. Instead of a computationally challenging sub-step of ALS and GD, we implement the algorithm on parallel machines by using only two sparse matrix-vector products. Not only is the algorithm scalable but it is also on average 4 to 10 times faster than competing algorithms on various data sets. Next, we discuss our results of non-negative matrix factorization for hyperspectral image data in the presence of noise. We introduce a spectral total variation regularization and derive four variants of the alternating direction method of multiplier algorithm. While all four methods belong to the same family of algorithms, some perform better than others. Thus, we compare the algorithms using stimulated Raman spectroscopic image will be demonstrated.

For deep neural network models, we focus on its application to image denoising. We first demonstrate how an optimal procedure leveraging deep neural networks and convex optimization can combine a given set of denoisers to produce an overall better result. The proposed framework estimates the mean squared error (MSE) of individual denoised outputs using a deep neural network; optimally combines the denoised outputs via convex optimization; and recovers lost details of the combined images using another deep neural network. The framework consistently improves denoising performance for both deterministic denoisers and neural network denoisers. Next, we apply the deep neural network to solve the image reconstruction issues of the Quanta Image Sensor (QIS), which is a single-photon image sensor that oversamples the light field to generate binary measures.

Share

COinS