Steered Mixture of Experts Regression for Image Denoising with Multi-Model-Inference

In this paper we introduce a novel block-based regression strategy for image denoising based on edge-aware Steered-Mixture-of-Experts (SMoE) models. SMoEs provide very sparse image representations, able to model sharp edges as well as smooth transitions in images efficiently with few parameters. A multi-model inference strategy is developed that improves significantly the denoising capacity of single SMoE models. We show that the important edge reconstruction properties of SMoEs are well preserved, even when many models are fused under severe noise. We investigate model-inference from local neighborhood blocks as well as from distant blocks using block-matching as in BM3D. Our initial results indicate that SMoE multi-model regression can provide promising results compared to state-of-the-art BM3D with excellent edge quality.