A Multi-class Magnitude Classifying Sparse Source Model for Compressible Sources

Source modelling is a gateway to the fascinating world of source coding. Many real-world sources are sparse or have a sparse representation. According to this fact, this work has focused on providing a new model to represent real-world non-strictly sparse (compressible) sources. To this aim, a novel model has been evolved from a simple sparse binary source to reflect the characteristics of compressible sources. The model is capable to represents real-world compressible sources by classifying samples into different classes based on their magnitudes. The model parameters are estimated using an innovative approach, a combination of a clustering technique and the binary genetic algorithm. The ability of the new approach has been assessed in modeling DCT coefficients of still images and video sequences. The proposed model also inspires an efficient coding approach to compress a wide range of sources including compressible sources. Comparison with classical well-known distributions including Laplace, Cauchy, and generalized Gaussian distribution and also with the most recent Noisy BG model reveals the capabilities of the proposed model in describing the characteristics of sparse sources. The numerical results based on the “chi-square goodness of fit” show that the proposed model provides a better fit to reflect the statistical characteristics of compressible sources.