Single-Pixel Image Reconstruction Based on Compressive Sensing and Deep Learning

This thesis seeks to improve the performance of single-pixel imaging. For data sampling, block compressive sensing (BCS) is introduced as an alternative to the conventional full-image compressive sensing. Its performance in simulation and on empirical data justified its use, besides the fact that it uses less storage and memory. Subsequently, a deep-learning model is proposed as the image reconstruction algorithm. Experiments show that the model is capable of reconstructing images obtained from an SPI setup while being priorly trained on natural images. This opens up opportunities for pretrained deep-learning models for BCS reconstruction of images from various domains.