Improving Sparse Dictionary Learning with Rejection of Fully Exposed Patches

Sparse dictionary learning is widely used in image processing tasks such as denoising, classification and image enhancement. However, learning features from X-ray images can be challenging since some parts of the image contains no or very little clinical information which makes them totally exposed to X-rays. These areas are irrelevant for medical content reconstruction and waste the space in the dictionary. Moreover, they slow down the convergence of the learning methods. In this paper, we present a prior knowledge about these fully exposed areas based on their physical distribution and propose a specialized dictionary learning algorithm to improve convergence speed and the quality of the learned dictionary while keeping the same quality of reconstruction. We compare our solution to a variance based approach and show that our solution is robust to the X-ray dose and then more efficient than the variance-based approach.