Fig. 5: Comparisons among BCNN, CNN, End-To-End BCNN and CNN in the experimental SPI trained with the MNIST database.
From: Approximating the uncertainty of deep learning reconstruction predictions in single-pixel imaging
a A representative ground-truth image in the testing dataset, input images to the BCNN and CNN calculated from the LSQR-approximated inverse model matrix, 1D raw measurement data as the input to End-To-End BCNN and End-To-End CNN, and the predictions from BCNN, CNN, End-To-End BCNN and End-To-End CNN at the 16× and 64× compression ratios. b A ground-truth image out of the testing dataset in the MNIST database, input images to the BCNN and CNN calculated from the LSQR-approximated inverse model matrix, 1D raw measurement data as the input to End-To-End BCNN and End-To-End CNN, and the predictions from BCNN, CNN, End-To-End BCNN and End-To-End CNN at the 16× and 64× compression ratios. c The MAEs of the predicted images in BCNN, CNN, End-To-End BCNN and End-To-End CNN at the two compression ratios. d The SSIMs of the predicted images in BCNN, CNN, End-To-End BCNN and End-To-End CNN at the two compression ratios. e The correlation coefficient, R, between the predicted uncertainty and the true absolute error of each pixel the predicted images reconstructed in BCNN and End-To-End BCNN at the two compression ratios. The error bars represent the standard deviation of the corresponding parameters from 100 testing images.
