Deep Learning Based Single Pixel Imaging Using Coarse-to-fine Sampling

Image quality and time efficiency are the primary concerns in single pixel imaging (SPI) system. In general, one can increase the number of measurements to improve the image quality, but this will overloads the acquisition and reconstruction process on the other hand. The improvement should not only address the image quality issue, but also needs to consider the efficiency. Therefore, this paper proposes a deep learning based SPI using coarse-to-fine sampling scheme. Benefits from the efficiency of deep learning reconstruction, the proposed method progressively samples and reconstructs a better image until a specific criterion is fulfilled. The results show that coarse-to-fine sampling consistently outperforms the uniform sampling in terms of image quality. At the same time, efficient image computation is achieved by the deep learning GAN based reconstruction. In conclusion, the proposed method is proven as a feasible solution to optimise the trade-off between image quality and computational load.