Combining Classic Methods with Deep Learning Priors for Lightweight Visual Denoising

Abstract

Image and video denoising is a fundamental problem in signal processing, with applications spanning from video compression and smartphone image pipelines to cinema post-production and medical imaging. While traditional denoising techniques rely on carefully crafted priors to model natural image properties, modern approaches increasingly employ deep-learning-based models that leverage large-scale data to achieve state-of-the-art results. However, the trend toward large and computationally expensive architectures, such as transformers, raises significant challenges in terms of practicality and efficiency.

This thesis explores innovative approaches to denoising by combining classical signal processing techniques with modern neural networks to create lightweight, high-performance denoisers. We first develop a set of guidelines for designing compact, real-noise image denoisers, focusing on training strategies, architectural considerations, and the effective use of synthetic-real noise datasets. Building on these principles, we present a hybrid Wiener-CNN image denoising framework that combines the classic Wiener filter with small auxiliary neural networks, achieving comparable performance to popular denoisers like DnCNN while maintaining an order-of-magnitude reduction in parameters.

Extending this hybrid approach to video denoising, we design a 4D Wiener-CNN framework capable of leveraging temporal and spatial frequency information. This model outperforms state-of-the-art transformers in high-noise tasks and offers significantly faster denoising times, demonstrating the potential of lightweight architectures in video restoration.

Finally, we explore the use of implicit natural image priors modelled by denoisers to solve linear inverse imaging problems, building on the framework proposed by Kadkhodaie and Simoncelli. By fine-tuning the denoiser within the iterative process, we improve its performance across multiple tasks, including inpainting, super-resolution, and compressed sensing. These findings highlight the broader potential of integrating classical and modern methods to address longstanding challenges in denoising, offering a path toward efficient and adaptable solutions.