Robust Dictionary Learning by Error Source Decomposition
Sumedh Chandaluri · 2018
Sparsity models have as of late, indicated incredible guarantee in numerous vision assignments. Utilizing a Learned Dictionary in sparsity models can as a rule beat predefined bases in clean information. Practically, both Training and Testing information might be defiled and contain noises and outliers. However, many methods are available for Face Recognition to deal with deformed data and acquired Super results in Testing Phase, but how to deal with defilement in Training stage still remains an exceptionally troublesome issue. As opposed to most existing strategies that take in the word reference from clean information, this thesis is focused at taking care of noises and outliers in training phase also. We propose a general strategy to break down the reconstructive residuals into two parts: a Non-Sparse segment for small universal noises and Sparse Segment for Large outliers. Analyses on Synthetic data and in addition Real applications have demonstrated agreeable execution of this new hearty word Dictionary learning approach.