Core Concepts
The author proposes an unsupervised bilevel learning strategy based on residual whiteness to estimate regularization parameters in imaging inverse problems, showing comparable results to standard metrics without requiring ground-truth data.
Abstract
The content discusses an unsupervised bilevel optimization strategy for learning regularization parameters in imaging inverse problems. It introduces a novel approach that optimizes the whiteness of the residual between observed data and the observation model without needing ground-truth data. The proposed method is validated on Total Variation-regularized image deconvolution problems, demonstrating close estimates to mean-square error oracle and discrepancy-based principles. Variational methods are highlighted as a reference paradigm for stabilizing unstable inversion processes by minimizing energy functionals encoding prior information on images and noise statistics. The article delves into variational terms, optimization problems, regularization choices, and parameter estimation challenges. It explores classical approaches like cross-validation and Morozov-type methods for estimating optimal hyperparameters. Bilevel learning is presented as a powerful paradigm for hyperparameter estimation, optimizing quality measures assessing solution goodness. The use of supervised, semi-supervised, and unsupervised metrics like MSE, Gaussianity loss, and residual whiteness principle is discussed in detail. Optimization algorithms like Nesterov's accelerated gradient-descent and Gauss-Newton solvers are explained for solving lower-level and upper-level problems iteratively. Experimental results comparing different quality metrics on image deconvolution tasks with various blur types and noise levels are provided along with numerical evaluations and visual reconstructions.
Stats
"We consider an unsupervised bilevel optimization strategy for learning regularization parameters."
"The proposed quality metric provides estimates close to the mean-square error oracle."
"For each test image and BSNR value we evaluated the quality of the image reconstructed by bilevel optimization."
Quotes
"Optimality of bλ is assessed by maximizing the whiteness of the residual between observations y and the observation model."
"Our results suggest that such quality measure performs as well as standard MSE-based alternatives."