New sparse representation methods; application to image compression and indexing
Year: 2010
Type: dissertation
Abstract: A new dictionary structure is introduced called an Iteration-Tuned Dictionary (ITD). ITDs are layered structures containing a set of candidate dictionaries in each layer. ITD-based iterative pursuit decompositions are carried out using, at each iteration i, one of the candidates from the i-th layer. A general ITD framework is proposed as well as a tree-structured variant called the Tree-Structured Iteration-Tuned Dictionary (TSITD) and a constrained tree-structured variant called the Iteration-Tuned and Aligned Dictionary (ITAD). These structures are shown to outperform various state-of-the-art reference algorithms in their ability to approximate a dataset sparsely, and in the applications of image denoising and image compression. The ITAD scheme, in particular, is used to develop an image codec that outperforms JPEG2000. An approximate vector search method is also introduced which uses sparse representations to carry out low-complexity approximate nearest-neighbor image searches. The approach addresses the related instability of the sparse support when the image patch is subject to weak affine transformations. In developing this new approach, a new data conditioning scheme is introduced that succeeds in better distributing data on the unit sphere while preserving relative angles. It is shown that this new approach improves the complexity/performance tradeoff of approximate searches based on sparse representations.
Source: HAL
Author: Joaquin Zepeda Salvatierra
Language: English
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W625249714