The article presents a novel training strategy for deep denoisers in signal and image processing. The strategy enforces a weaker constraint on the deep denoiser, termed pseudo-contractiveness, which improves recovery performance. The study explores the relationship between different denoiser assumptions through spectral analysis of the Jacobian matrix. Algorithms based on gradient descent and Ishikawa process are derived, which theoretically converge strongly to a fixed point. The pseudo-contractive denoiser demonstrates superior performance in experiments, offering competitive visual effects and quantitative values.
Publication date: 8 Feb 2024
Project Page: https://arxiv.org/abs/2402.05637v1
Paper: https://arxiv.org/pdf/2402.05637
