A number of image-processing problems can be formulated as optimizationproblems. The objective function typically contains several terms specificallydesigned for different purposes. Parameters in front of these terms are used tocontrol the relative weights among them. It is of critical importance to tunethese parameters, as quality of the solution depends on their values. Tuningparameter is a relatively straightforward task for a human, as one canintelligently determine the direction of parameter adjustment based on thesolution quality. Yet manual parameter tuning is not only tedious in manycases, but becomes impractical when a number of parameters exist in a problem.Aiming at solving this problem, this paper proposes an approach that employsdeep reinforcement learning to train a system that can automatically adjustparameters in a human-like manner. We demonstrate our idea in an exampleproblem of optimization-based iterative CT reconstruction with a pixel-wisetotal-variation regularization term. We set up a parameter tuning policynetwork (PTPN), which maps an CT image patch to an output that specifies thedirection and amplitude by which the parameter at the patch center is adjusted.We train the PTPN via an end-to-end reinforcement learning procedure. Wedemonstrate that under the guidance of the trained PTPN for parameter tuning ateach pixel, reconstructed CT images attain quality similar or better than inthose reconstructed with manually tuned parameters.
展开▼
