On the use of deep learning for single-pixel imaging

We apply deep learning (DL) to counter three key problems which may occur in single-pixel imaging (SPI) namely noise, appearance of ringing or pixelated artifacts due to undersampling, and effects of projector lens aberration or defocusing. We employ a multi-scale mapping based deep convolutional neural network (DCNN) architecture to rectify undesirable effects in a 96x96 target reconstruction produced by environmental or system conditions, and optical anomalies. We train the proposed DCNN on augmented experimental data as well as simulation data to achieve robust experimental performance. Experimental results on real targets (2D and 3D) demonstrate the superior performance of the proposed method compared to conventional SPI.