Optoelectronic Attention-Based Distillation Network for Color Image Denoising

Abstract

Although deep learning-based image denoising achieves excellent performance, its massive parameters and computational complexity hinder deployment in low-power edge systems. To address these challenges, this letter proposes an Optoelectronic Attention Denoising Network (OEADNet) for color images. It consists of a single-layer multi-kernel parallel metasurface optical front-end for feature extraction and a lightweight Restormer-style electronic back-end for restoration. By combining a Restormer teacher network with spatial domain result distillation, it achieves end-to-end collaborative optimization. The model requires 0.364 M parameters and 21.72 G FLOPs, accounting for only 1.39% and 7.70% of the teacher network. Numerical experiments demonstrate stable restoration across various Gaussian noise levels. At a noise level of $\boldsymbol {\sigma } =50$ for 8-bit images, OEADNet improves upon a same-scale pure electronic baseline by up to 1.41 dB, and surpasses denoising convolutional neural networks (DnCNNs) by 0.77 dB. Furthermore, it extends optical neural networks from grayscale scenes to color denoising, providing a design that balances performance and physical realizability for resource-constrained systems.