Few-shot trained single-lens diffractive neural network for all-optical image denoising

Image denoising technology has undergone extensive research over the past few decades. Pioneering studies have developed analog diffractive image denoisers that leverages cascaded diffractive neural network (DNN) architecture to remove noise at a speed of light. However, existing all-optical DNN denoisers require large amounts of training data and exhibit poor generalization. Besides, such denoisers employs multi-layer structure to enhance performance, while bringing difficulty in aligning and limitation in diffraction efficiency. To address these challenges, we propose and experimentally demonstrate what we believe to be a novel few-shot trained single-lens DNN (SL-DNN) for image denoising. By incorporating the DNN into a single-lens imaging system which introduces the spatial frequency filtering priors, the feature representation capability of SL-DNN is effectively enhanced. As a proof of concept, SL-DNN architecture adopts only one diffractive layer to achieve better image denoising performance compared with regular five-layer DNN, while avoiding alignment difficulties and diffraction efficiency losses. Moreover, SL-DNN only requires two orders of magnitude fewer training data, while achieving better generalization. The superior generalization of the SL-DNN model is experimentally verified on different datasets such as Quick Draw and ECSSD, although being trained using few-shot data from a single dataset. The proposed SL-DNN method provides a physics-inspired strategy for designing visual processors for specific tasks, including image display and projection applications.