Research

 

Overview

Finding relevant information in most natural signals is like looking for a needle in a haystack; as such most models of signals (e.g. those that model signals as members of a vector space) exhibit too many degrees of freedom. To overcome this challenge, sparse models of signals have been introduced ranging from those that limit the number of nonzeroes in a vector to manifold models and low-rank matrices.

The Sparse Signal Processing Group (SSPG) at DML is focused on the application of sparse signal representations and compressed sensing theory to challenging signal processing tasks such as image categorization, remote sensing, sampling of complex networks, etc. Our current research is focused on the following projects:

Capturing Dependencies of Dictionary Atoms for Image Classification

Spatial-Aware Dictionary Learning for Hyperspectral Image Classification 

Spatially Weighted Sparse Coding for Hyperspectral Image Classification

Semi Spatio-Temporal fMRI Brain Decoding

Visual Tracking via Joint Sparse Representation

 
 
 

The purpose of remote sensing is to analyze data obtained by satellite images from ground (of earth or any other planet). Among its applications are monitoring urban areas and forecasting the location of fire and the ocurrence of flood. The data used here are multispectral images, which are captured from each point in different bands, providing lots of information for precise analysis.

One of the issues in this field is labeling different areas according to their texture. Since labeling all the training data is mostly impractical, we are encouraged to use semi-supervised learning approaches.

In this project we assume that the data lie on a manifold and the labeling function is smooth on the neighborhood graph (as an estimation of the manifold). We pursue semi-supervised learning approches for the task at hand.

A new aproach to graph construction and label inference problems are based on sparse representation of data i.e. finding appropriate bases to represent data sparsely.

People involved: Ahmad Khajenezhad, Mohammad H. Rohban, Ali Soltani-Farani, Hamid R. Rabiee.

Sparse Signal Processing Group