Single-pixel imaging with optical phase arrays: a machine-learning-assisted approach

Speckle patterns used for imaging are generated with a single pixel detector employing optical phased arrays (OPAs). The OPA modulates light, which then travels through a diffraction grating, creating speckle patterns via various light paths. Once the target is illuminated, these patterns are captured by a bucket detector in single-pixel imaging, and the image is reconstructed using Compressed Sensing (CS) techniques.1 In this work, we propose a Machine Learning (ML)2 based approach to enhance the quality of CS reconstructions by comparing two popular architectures for image denoising: a basic Convolutional Neural Network (CNN) and a U-Net. A dataset of paired noisy CS target image and clean images was generated and used to train and evaluate both models. Quantitative results indicate that the U-Net significantly outperforms the CNN in terms of Mean Squared Error, Structural Similarity Index, and Peak Signal-to-Noise Ratio, demonstrating its effectiveness in maintaining structural information while reducing noise. This study opens the door the potential of ML in boosting reconstruction fidelity for single pixel imaging systems based on OPAs