Block dictionary learning with l0 regularization and its application in image denoising

Dictionary learning is aimed to learn a set of basic elements termed as atoms from a given training set, and these atoms form a dictionary. In this paper, we propose a block dictionary learning algorithm, called Mini-batch K-sparse Dictionary Learning (MKDL), by directly optimizing a K-sparse problem under the mini-batch setting. At each iteration of MKDL, only small-batch training samples are used to sparse coding and dictionary update stages. Particularly, iterative hard thresholding and projected gradient descent schemes are employed to optimize the two above-mentioned stages, respectively. Preliminary results on image denoising have much better performance than previous dictionary learning algorithms, which validates the effectiveness of our approach in convergence speed and denoising quality.