Tomographic Image Reconstruction using Dictionary Priors

We describe and examine a framework for tomographic image reconstruction where prior knowledge about the solution is available in the form of training images. We first construct a dictionary that contains prototype elements from these images. Then by using the dictionary as a prior to regularize the inverse problem, and looking for a solution with a sparse representation in the dictionary, we formulate the reconstruction problem in a convex optimization framework. Our computational experiments clarify the choice and interplay of the model parameters and the regularization parameters, and they show that in few-projection settings we are able to produce better images with more structural features than the total variation approach.