Adaptive Regularization in Convex Composite Optimization for Variational Imaging Problems

We propose an adaptive parameter balancing scheme in a variational framework where a convex composite energy functional that consists of data fidelity and regularization is optimized. In our adaptive parameter balancing, the relative weight is assigned to each term of the energy for indicating its significance to the total energy, and is automatically determined based on the data fidelity measuring how well the data fits the model at each energy optimization step. The adaptive balancing parameter is designed to locally control the regularity by taking into consideration the residual at each point without introducing any a-priori knowledge. We demonstrate that our adaptive balancing parameter is effective to improve the quality of the solution by determining the degree of regularity based on the local residual in the application of imaging problems. We apply our adaptive parameter balancing scheme to the classical imaging problems that are denoising, segmentation and motion estimation, and provide their optimization algorithms based on the alternating direction method of multipliers (ADMM) method. The robustness and effectiveness of our adaptive parameter balancing is demonstrated through experimental results presenting that the qualitative and quantitative evaluation results of each imaging task with an adaptive balancing parameter is superior to the results with a static one. The desired properties, robustness and effectiveness, of the parameter selection in a variational framework for imaging problems are achieved by merely replacing the static balancing parameter with our adaptive one.