A temporal correlation superresolution image is based on the variance of the recorded photon time trace, while its resolution is higher than that of the complementary intensity image, it is noisier. Both images, the intensity and correlation based, are fed into a deep convolutional neural network (CNN), which produces an image that is optimized to have higher resolution than the intensity image and less noise than the correlation image. The image then passes through separate linear networks that mimic the physical blurring of the imaging setup. Preliminary experimental results show similar resolution to the experimental superresolution image with less noise.
|