Inference with l0-norm-based Sparsity Prior on Discrete Framework

We present a new penalizing scheme for a recently introduced prior model [8] on discrete frameworks. The model convincingly assumes that the optimal solutions for the frameworks possess sparse representation on certain transform domains, and applies this sparsity assumption as a prior information for inference problems. Promoting the sparsity, we proposes to penalize l0-norm of coefficient vector of the transform bases, instead of l1-norm employed in that recent work. Experiments compare the proposed prior with previous ones and show enhanced performance, both in qualitative and quantitative manner.