Fig. 5: Simulations to quantify RDM performance.

a, Error maps showing the absolute difference between ring convolution/standard convolution and the ‘true blur’ (produced by manually superimposing every PSF at every pixel). When off-axis aberrations are small (top row), both forward models are accurate. When aberrations are large (bottom row), convolution becomes noticeably worse, yet ring convolution remains accurate. b, Forward model mean-squared error (MSE) as a function of off-axis aberration magnitude (for the image above). c, Runtime of each method as a function of size of the image (in pixels), averaged over n = 50 trials. d, Seidel coefficients are fit to a noisy image of randomly placed PSFs and then used to generate PSFs at any location. e, Seidel fit error at every iteration of the fitting algorithm. Each purple line is one of n = 50 trials, each with a different, random set of underlying Seidel coefficients. The red dashed line is the per-iteration median. f, Average MSE of the fitted Seidel coefficients plotted against the SNR of the calibration image (with standard deviation shown by error bars) over n = 50 random trials. Some example calibration PSFs are shown. g, Deblurring results on noisy images from the CARE dataset, with PSNR values above each method. h, Zoom-ins of an off-axis patch in each deblurred image; ring deconvolution and DeepRD have the highest quality. i, Average accuracy in PSNR versus runtime of each method over n = 28 true blurred images using unseen coefficients. The number of model parameters is written below each circle and determines its size.