Robust Non-convex Model-Based Approach for Deep Learning-Based Image Processing

Total p-norm Variation (TpV) is a well-established technique in image processing, used to denoise and preserve edges. However, the related non-convex minimization is still a challenging task in optimization, both for the computational cost and for convergence guarantees toward a good local minimum. We propose a framework, called (TpV) , embedding a convolutional neural network to speed up the iterative reconstruction while preserving converging features. The resulting hybrid method is robust and accurate. Verifications and comparisons illustrate that the proposed method is effective and promising.