Blind Source Separation (BSS) is a problem where signals (audio/images) are separated from given mixtures. The classical BSS problem is explained well using the "cocktail party scenario"
Abstract
Blind Source Separation (BSS) is a problem where signals (audio/images) are separated from given mixtures. The classical BSS problem is explained well using the “cocktail party scenario”: A room with several speakers talking simultaneously has several microphones placed in it. Each microphone records a different superposition of the speaker voices. In this scenario, BSS will try to separate the recordings so that we can hear each speaker without the other signals interfering.
In the classical BSS problem we assume independence of time/space. In reality this is not accurate. Taking the cocktail party scenario- if the speakers walk around the room the problem is time depending. Therefore we will examine a time/space depending problem.
This project will deal with BSS of images. The separation of the original signals (images) will assume that they undergo any affine transform and any spatial dependent amplification transform before mixing. No other knowledge of the original images is assumed, apart from being natural images.
We will attempt to solve the BSS problem by estimating the relative transforms between the mixtures. This will be done using sparse component analysis (SCA): the mixtures will be converted into sparse representation. By comparing sparse representations we will estimate the relative transforms.
After estimating the transformations, we will attempt to solve an inversion problem to reconstruct the original images. This will be done using iterative algorithms for solving optimization problems.
Flowchart
Project characteristics:
- Space variant Non-instantaneous BSS problem
- Separating two images
- Transforms are combination of Affine & Spatial gain
- No prior knowledge about the sources except that they are natural images
- Separation will be done using Sparse Component Analysis
Results
