Efficient half-quadratic regularization with granularity control

摘要

In the last decade, several authors have proposed edge preserving regularization methods for solving ill posed problems in early vision. These techniques are based on potentials derived from robust M-Estimators. They are capable of detecting outliers in the data, finding the significant borders in noisy images and performing edge-preserving restorations. These methods, however, have some problems: they are computationally expensive, and often produce solutions which are either too smooth or too granular (with borders around small regions). In this paper, we present a new class of potentials that permits to separate robustness and granularity control, producing better results than the classical ones in both scalar- and vector-valued images. We also present a new fast, memory-limited minimization algorithm.