Low bit-rate SNR scalable video coding based on overcomplete dictionary learning and sparse representation

• Published in: Multidimensional systems and signal processing Vol. 31; no. 2; pp. 465 - 489
• Main Authors: Irannejad, Maziar, Mahdavi-Nasab, Homayoun
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.04.2020; Springer Nature B.V
• Subjects: Algorithms; Artificial Intelligence; Circuits and Systems; Coding; Dictionaries; Discrete cosine transform; Electrical Engineering; Engineering; Heterogeneity; Image compression; Learning; Representations; Signal to noise ratio; Signal,Image and Speech Processing; Singular value decomposition; Video compression; Video signals; Scalable video coding; Quantization; Sparse coding; SNR; Dictionary learning
• ISSN: 0923-6082; 1573-0824
• DOI: 10.1007/s11045-019-00671-6

In the current decade, scalability has been developed in video coding (VC) schemes to reply end-user demands and heterogeneity of networks. In this paper, a low bit-rate signal-to-noise ratio (SNR) scalable VC based on dictionary learning (DL) and sparse representation is proposed. A notable feature of SNR scalability compared to spatial and temporal versions is that there are not any limitations in the number of enhancement layers, making it more applicable to adapt to different conditions. In this research, unlike traditional VC in which the discrete cosine transform (DCT) coefficients of video signals are quantized to obtain different SNR qualities, sparse codes are applied. Sparse coding is done over trained overcomplete dictionaries, for which three different DL algorithms, namely MOD, K-SVD, and RLS-DLA, are utilized and compared. The dictionaries are trained over the DCT domain of general natural images, to achieve higher compression and prevent blocking artifacts. The results of the proposed method are compared with non-scalable coding based on DL, and scalable and non-scalable coding schemes based on complete DCT dictionary employed in traditional VC standards such as MPEG.X and H.26X. The results show that, although video scalability naturally decreases the quality compared to non-scalable coding, the proposed scheme presents superior subjective and rate–distortion performance compared to non-scalable and scalable VC based on the traditional DCT quantization. Moreover, among the three DL methods applied, RLS-DLA achieves superior results both for non-scalable and scalable VC.