טכניון מכון טכנולוגי לישראל
הטכניון מכון טכנולוגי לישראל - בית הספר ללימודי מוסמכים  
M.Sc Thesis
M.Sc StudentOri Bryt
SubjectFace Image Compression Using Sparse and Redundunt
Representations and the K-SVD Algorithm
DepartmentDepartment of Electrical Engineering
Supervisor Full Professor Elad Michael
Full Thesis textFull thesis text - English Version


Abstract

The use of sparse and redundant representations in signal and image processing has been gradually increasing over the past decade. Obtaining an overcomplete dictionary from a set of signals allows us to represent them as sparse linear combinations of dictionary atoms, when pursuit algorithms are used for signal decomposition. One of the most intuitive ways to obtain such a dictionary is training it from a database of signals, and several algorithms have been proposed to handle this task. A notable work in this field introduced the K-SVD algorithm, which is an efficient method for such training. In this research we propose a novel method for compressing facial images, based on the K-SVD algorithm. Facial images are an important class of images, used in databases for large organizations and government institutions, and compressing them to low bit-rates while maintaining image quality is of great value. We train K-SVD dictionaries for predefined image patches, and compress every new image according to these dictionaries. The encoding is based on the decomposition of each image patch using sparse coding over the relevant trained dictionary, and the decoding is a simple reconstruction of the patches by a linear combination of the relevant dictionary atoms. We examine several parameters that affect the performance of our method, such as dictionary redundancy and patch sizes. An essential pre-process stage is an image alignment procedure, where several facial features are detected and geometrically warped into a canonical spatial location. This increases the redundancy in the images and allows each dictionary to be trained on much similar data. Another significant part of the algorithm is a post-process of image deblocking. Since the encoding is done on non-overlapping patches, a visually disturbing artifact of blockiness appears in the reconstructed images. Our goal is to suppress this artifact, and we do so using a simple linear deblocking technique, which is based on local image filters. We analyze the performance of our compression method, present results and compare it to several competing compression techniques.