Home


A Penalty Function Promoting Sparsity Within and Across Groups


Software

The file SWAG.zip contains Matlab codes used in the experiments from the manuscript "A Penalty Function Promoting Sparsity Within and Across Groups", by I. Bayram and S. Bulek, 2016.


Audio Denoising Experiment

Below are some results from the audio denoising experiment from the manuscript. In the manuscript, due to lack of space, we show the results only for input SNR = 5 dB. Here we also provide the sound files and spectrograms for when input SNR = 0 dB.

The original signal is speech but it is degraded by additive ambient noise.


Noisy Observation

SNR = 5 dB

Here are the reconstructions obtained by the methods described in the manuscript.

l1 Norm Regularization

Output SNR = 6.42 dB

E-Lasso Norm Regularization

Output SNR = 6.54 dB

Mixed (l2,1) Norm Regularization

Output SNR = 6.09 dB

Proposed Penalty

Output SNR = 6.67 dB

Hybrid (proposed + l2,1) Penalty

Output SNR = 7.55 dB

Below are the spectrograms of the signals


Below are the noisy signal and the denoised ones using different methods for input SNR = 0 dB.

Noisy Observation

SNR = 0 dB

l1 Norm Regularization

Output SNR = 2.48 dB

E-Lasso Norm Regularization

Output SNR = 2.63 dB

Mixed (l2,1) Norm Regularization

Output SNR = 2.38 dB

Proposed Penalty

Output SNR = 2.64 dB

Hybrid (proposed + l2,1) Penalty

Output SNR = 3.86 dB