Steered mixture-of-experts autoencoder design for real-time image modelling and denoising

Abstract

Research in the past years introduced Steered Mixture-of-Experts (SMoE) as a framework to form sparse, edge-aware models for 2D- and higher dimensional pixel data, applicable to compression, denoising, and beyond, and capable to compete with state-of-the-art compression methods. To circumvent the computationally demanding, iterative optimization method used in prior works an autoencoder design is introduced that reduces the run-time drastically while simultaneously improving reconstruction quality for block-based SMoE approaches. Coupling a deep encoder network with a shallow, parameter-free SMoE decoder enforces an efficent and explainable latent representation. Our initial work on the autoencoder design presented a simple model, with limited applicability to compression and beyond. In this paper, we build on the foundation of the first autoencoder design and improve the reconstruction quality by expanding it to models of higher complexity and different block sizes. Furthermore, we improve the noise robustness of the autoencoder for SMoE denoising applications. Our results reveal that the newly adapted autoencoders allow ultra-fast estimation of parameters for complex SMoE models with excellent reconstruction quality, both for noise free input and under severe noise. This enables the SMoE image model framework for a wide range of image processing applications, including compression, noise reduction, and super-resolution.

Denoising OCT images using steered mixture of experts with multi-model inference

“…2 demonstrates the efficiency of the proposed training strategy with BM-SMoE-AE model. The encoder utilizing composite loss outperformed the previous model,10 achieving an average improvement of approximately 0.7 dB in PSNR and 0.02 in SSIM.…”

“…As the objective function for the SMoE Gating Network, we utilized a composite loss 15 comprising Mean Squared Error (MSE) and the Structural Similarity Index (SSIM), as depicted in Equation 10, differing from the approach in recent work. 10 However, in our previous study, 16 we employed SSIM as the loss function in GD for global optimization in image compression task.…”