טכניון מכון טכנולוגי לישראל
הטכניון מכון טכנולוגי לישראל - בית הספר ללימודי מוסמכים  
M.Sc Thesis
M.Sc StudentGrosberg Unger Hilit
SubjectInverse Problems in Neural and Radiometric Systems
DepartmentDepartment of Electrical Engineering
Supervisor Professor Emeritus Zeevi Yehoshua
Full Thesis textFull thesis text - English Version


Abstract

Blind Source Separation (BSS) is a powerful approach to a wide range of unmixing problems. We adopt the sparsification method for the solution of such problems. This approach lends itself to Sparse Component Analysis (SCA). The SCA is based on projection of the mixtures' data onto a scatter space of sparse representations, wherein the unknown mixing matrix can be estimated by simple geometrical means. We extend the SCA algorithm to the spatio-temporal domain and apply it in unmixing of dynamic reflections superimposed on a target.


We present initial results of recovering an estimated "clear-day" image of an acquired weather-degraded image. Motivated by BSS, we investigate the recovery from only one acquired color image by combining and extending previous related studies. We incorporate prior knowledge about the scene, such as colors or depths of objects. These priors are used in the process of recovering a partial depth map of the scene. A clear-day scene is recovered from the degraded one by interpolating this sparse map.


Lastly, we investigate also the separation of massive neural activity into simpler neural components by BSS and similar decomposition methods. Unlike the separation of reflections, the activity of neurons is inherently non-linear and, in addition, the number of sources is a priori unknown. We show that the most suitable method for the extraction of such sources ("cliques") from their mixtures is the Non-negative Matrix Factorization (NMF) algorithm. We search for good estimators of the number of sources for the case of neural data. We compare the performance of two previous methods found in the literature and propose two new methods that demonstrate better results on simulated  neural-like data.