Accelerated stochastic iterative algorithms for image reconstruction via enhanced Regularization-by-Denoising
Abstract
Inverse imaging problems such as denoising and deblurring aim to recover a clean image from corrupted measurements by incorporating prior knowledge. Classical regularizers like Tikhonov and total variation are mathematically convenient, yet they often oversimplify the complex structure of natural images. Regularization by Denoising (RED) uses a learned denoiser and yields strong theoretical guarantees, but typical implementations rely on full-batch updates that become computationally costly for large-scale problems. In this paper, we propose two accelerated stochastic formulations of RED that combine RED’s fidelity with the efficiency of stochastic optimization. In the direct formulation, we develop a stochastic primal–dual fixed‑point algorithm that computes mini‑batch gradients, applies variance reduction and uses a relaxation parameter for inertial updates. In the constrained formulation (RED‑PRO), we introduce a complementary algorithm that employs adaptive dual scaling and projection onto the denoiser’s fixed-point set to reduce denoiser calls. Both formulations dispense with strong convexity in favour of Polyak–Łojasiewicz, quadratic-growth and metric subregularity conditions, enabling linear convergence under realistic assumptions. A unified analysis proves convergence, and extensive experiments show that both stochastic RED formulations match or surpass state-of-the-art methods while significantly reducing computational cost and maintaining high PSNR and SSIM.