Adaptive structured block sparsity via dyadic partitioning

This paper proposes a novel method to adapt the block-sparsity structure to the observed noisy data. Towards this goal, the Stein risk estimator framework is exploited, and the block-sparsity is dyadically organized in a tree. The adaptation of the sparsity structure is obtained by finding the best recursive dyadic partition, whose terminal nodes (leaves) are the blocks, that minimizes a data-driven estimator of the risk. Our main contributions are (i) analytical expression of the risk; (ii) a novel estimator of the risk; (iii) a fast algorithm that yields the best partition. Numerical results on wavelet-domain denoising of synthetic and natural images illustrate the improvement brought by our adaptive approach.