Improving Pursuit Algorithms Using Stochastic Resonance
Sparse Representation Theory is a sub-field of signal processing that has led to cutting edge results in many applications such as denoising, deblurring, super resolution and many other inverse problems. Broadly speaking, this field puts forward a model that assumes that signals are originated from a sparse representation in terms of an over-complete dictionary. Thus, when a corrupted measurement is given, we seek to estimate its original, clean form by finding the best matched sparse representation of the given signal in the dictionary domain. This process is essentially a non-linear estimation solved by a pursuit or a sparse coding algorithm. The concept of Stochastic Resonance (SR) refers to the counter-intuitive idea of improving algorithms' performance by a deliberate noise contamination. In this work we develop novel techniques that apply SR for enhancement of the performance of known pursuit algorithms. We show that these methods provide an effective MMSE approximation and are capable of doing so for high-dimensional problems, for which no alternative exists.
READ FULL TEXT