Looking for optimal balance in Joint Denoising Demosaicing (JDD) using Deep Learning architectures

Abstract

[eng] —This thesis investigates the optimal balance in joint denoising and demosaicing (JDD) using deep learning architectures. Traditional image processing pipelines treat denoising and demosaicing as separate tasks, often resulting in suboptimal performance due to their interrelated nature. Recently, unified neural networks have emerged that perform these tasks jointly, improving results at the cost of losing control over the network’s behavior and increasing the number of parameters. We propose a solution that uses neural networks within an architecture allowing control over the system at intermediate stages, ensuring image-like output at each step. Specifically, our method integrates denoising both before and after demosaicing, utilizing multiple specialized networks to enhance image quality and reduce artifacts. We employ a pretraining strategy to optimize each network module individually before fine-tuning the entire system. A combined loss function is introduced for better control over the denoising process, and noise feedback mechanisms are explored by testing two different architectures that reintroduce noise at distinct pipeline stages. The model’s performance is evaluated against state-of-the-art methods on several well-regarded datasets, including McMaster, Kodak, BSD100, and Set14. Results show that our model consistently achieves higher CPSNR and MSSIM scores across various noise levels, demonstrating its superiority in noise removal and preserving image details. Additionally, we examine the impact of removing the raw denoiser and compare two noise feedback approaches, providing insights into the necessity of denoising at different stages. The findings contribute to developing more robust and efficient image processing techniques, with potential applications in consumer cameras and other imaging systems