Facial image compression using adaptive multiple dictionaries

In this paper a new image compression scheme using redundant dictionary and sparse coding is proposed. Unlike other sparse coding schemes which use just one dictionary we employ multiple specific dictionaries for compressing a class of facial images. The recursive least square dictionary learning algorithm, RLS-DLA, is used to design the adaptive dictionaries, each tuned to an interval of target compression rate. The evaluation of the presented method shows that in spite of being simple and fast, it outperforms modern standard compression techniques, specially the JPEG2000, by about 0.5 to 1.2 dB. This in turn, displays the effectiveness of the scheme compared to the state-of-the-art sparse coding schemes.