Image restoration from noisy and limited measurements with applications in 3D imaging

The recovery of image data from noisy and limited measurements is an important problem in image processing with many practical applications. Despite great improvements in imaging devices over the past few years, the need for a fast and robust recovery method is still essential, especially in fields such as medical imaging or remote sensing. These methods are also important for new imaging modalities where the quality of data is still limited due to current state of technology. This thesis investigates novel methods to recover signals and images from noisy or sparse measurements, in new imaging modalities, for practical 3D imaging applications. In particular, the following problems are considered. First, the Tree-based Orthogonal Matching Pursuit (TOMP) algorithm is proposed to recover sparse signals with tree structure. This is an improvement over the Orthogonal Matching Pursuit method with the incorporation of the sparse-tree prior on the data. A theoretical condition on the recovery performance as well as a detailed complexity analysis is derived. Extensive experiments are carried out to compare the proposed method with other state-of-the-art algorithms. Second, a new point clouds registration method is investigated and applied for 3D model reconstruction with a depth camera, which is a recently introduced device with many potential applications in 3D imaging and human-machine interaction. Currently, the depth camera is limited in resolution and suffers from complex types of noise. In the proposed method, the Implicit Moving Least Squares (IMLS) method is employed to derive a more robust registration method which can deal with noisy point clouds. Given a good registration, information from multiple depth images can be integrated together to help reduce the effects of noise and possibly increase the resolution. This method is essential to bring commodity depth cameras to new applications that demand accurate depth information. Third, a hybrid system which consists of a light-field camera and a depth camera rigidly attached together is proposed. The system can be applied for digital refocusing on an arbitrary surface and for recovering complex reflectance information of a surface. The light-field camera is a device that can sample the 4D spatio-angular light field and allows one to refocus the captured image digitally. Given light-field information, it is possible to rearrange the light rays appropriately to render novel views or to generate refocused photographs. In theory, it is possible to estimate the depth map from a light field. However, there is a trade-off between angular and spatial resolution in current designs of light-field cameras, which leads to low quality and resolution of the estimated depth map. Moreover, for advanced 3D imaging applications, it is important to have good quality geometric and radiometric information. Thus, a depth camera is attached to the light-field camera to achieve this goal. The calibration of the system is presented in detail. The proposed system is demonstrated to create a refocused image on an arbitrary surface. However, we believe that the proposed system has great potential in more advanced imaging applications.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Electrical & Computer Engr

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2015

Author dc:creator

• Bui, Huy Quang

Contributors dc:contributor

• Do, Minh N.
• Bresler, Yoram
• Hart, John
• Huang, Thomas
• Singer, Andrew

• compressed sensing
• tree-based
• Tree-based Orthogonal Matching Pursuit
• point-cloud registration
• Iterative Closest Point
• Signed Distance Registration
• light-field
• digital refocusing
• light-field camera
• depth camera

Statement dc:rights

• Copyright 2015 Huy Quang Bui

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/78597

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/78597