Towards high-resolution undersampled single-pixel imaging: A neural network perspective

This paper presents a novel single-pixel imaging (SPI) framework which can produce high-resolution target images with undersampling. Undersampling is used to work around the problem of long imaging time in SPI for real-time applications. However, the reconstruction from undersampled measurements suffers from noise, ringing or pixelated artifacts, and low resolution which complicates target recognition. To improve image quality, deep learning (DL) based approaches have been proposed but the improvement is merely based on noise and artifact removal. In order to improve image resolution, it is necessary to recover fine details from undersampled input which is very challenging due to absence of high-frequency information (during target reconstruction). To achieve this task, we propose to apply a DL model which learns to generate both low and high-frequency representations from an undersampled (10%) 96*96 input, and combines them to produce a high-quality (high-resolution) output. Experimental results show that the proposed model is robust against noise and frequency-based artifacts, and reconstructs high-quality targets by improving resolution (fine details).