Outlier-Robust Diffusion Solvers for Inverse Problems

Researchers have introduced a novel method to enhance the robustness of diffusion models (DMs) when solving inverse problems, particularly in the presence of outliers common in real-world measurements. While DM-based methods have shown remarkable performance recently, they often struggle with data anomalies. This new approach first refines measurements through explicit noise estimation to mitigate noise effects. It then formulates an iteratively reweighted least squares objective using Huber loss to effectively address outliers. To solve the resulting optimization problem, the authors propose a gradient descent method, further improved by employing the conjugate gradient method with an efficient update strategy to avoid sensitive learning rate tuning. Extensive experiments across multiple image datasets, covering both linear and nonlinear tasks under various conditions, demonstrate that this proposed method exhibits superior robustness to outliers. The results indicate that it outperforms recent DM-based methods in most scenarios, marking a significant advancement in computer vision and pattern recognition techniques for handling noisy and corrupted data in inverse problem solutions.