Author

Wentian Zhou

Date of Graduation

2017

Document Type

Dissertation

Degree Type

PhD

College

Statler College of Engineering and Mineral Resources

Department

Lane Department of Computer Science and Electrical Engineering

Committee Chair

Daryl Reynolds

Committee Co-Chair

Xin Li

Committee Member

Hailin Li

Committee Member

Natalia Schmid

Committee Member

Matthew Valenti

Abstract

Obtaining high-resolution images is a fundamental challenge for many vision related tasks. It is highly desirable to develop effective and efficient super-resolution algorithms to enhance the spatial resolution of acquired images. However, estimating high-resolution images is an ill-posed problem where a large set of possible solutions exist. To make this inverse problem more tractable, it is often necessary to constrain the targeted high-resolution image with a priori information (i.e., in model-based approaches) or to assume the availability of additional training data (i.e., in learning-based approaches). In this dissertation, we propose to (i) use simultaneous-sparse-coding and machine learning techniques to incorporate side information from a guided image and (ii) to study the role of data representations (e.g., spatial vs. frequency) in competing learning based approaches toward image super-resolution.;We first study the problem of depth map super-resolution, a special class of images captured by 3D depth sensing devices. The generated depth map is often in low resolution (LR) but paired with a high-resolution (HR) color image of the same scene. In one approach, simultaneous-sparse-coding (SSC) is employed to model the relationship between LR and HR depth maps. We extend the nonlocal similarity of the original SSC model from the guided color image to depth map in order to determine the unknown patch clusters. In another approach, we directly learn the relationship between LR depth map and HR color image using a recently developed deep neural network scheme.;The next problem in our study is the resolution-enhancement of natural images. Such a problem can be formulated in two different ways: interpolation and super-resolution (SR). We have developed a structural aware interpolation scheme capable of recognizing structural variations and inferring the unknown pixels based on the structural prior. Moreover, we propose a super-resolution scheme utilizing the local self-similarity property. Our approach overcomes the limitation of integer scaling factors and generates aesthetically pleasing SR results. Extensive evaluations have shown that the performance of the proposed methods achieve and exceed current state-of-the-art algorithms.;Overall, we propose to tackle three sub-problems of the image super-resolution problem from two different aspects: guided image and learning representation. To utilize the guided image, two competing approaches (model-based vs. learning based) are presented to incorporate the side information. Additionally, spatial domain learning and progressively high-frequency learning frameworks are compared to tackle the image interpolation and image SR problems respectively.

Share

COinS