We approach to the problem of inverse halftoning within the frameworks of Bayesian inference and compressed sensing, which is one of the most effective signal processing methods through sparse representation.
In this paper, we adopt the K-SVD dictionary for the sparse representation of an original image to be inferred, and develop our previous work with the DCT dictionary restricted to a small number of the slowest basis vectors. The K-SVD dictionary is known to have higher efficiency for sparse representation than the DCT one. Therefore, we can expect that it helps us overcome a heavily ill-posed property of the problem.
Numerical analysis confirms the effectiveness of our approach with the K-SVD dictionary, and makes clear the difference between the characteristics of the K-SVD dictionary and those of the restricted DCT one.