Optimal CNN templates for deconvolution

A simplified version of the gradient descent method is introduced as a straightforward way to find optimal 3x3 CNN templates for the inversion of known point spread functions (PSF). In practical applications the determination of this inverse is necessary to fulfil deconvolution tasks. The proposed method is much faster than the previously applied algorithms (like genetic algorithm) and still, in almost all practically important cases, it is convergent. Moreover, unlike a closed form method [1], it leads to 3x3 templates instead of 5x5 or bigger ones. In several important practical cases the PSF, which can be caused by motion, out of focus or the aberration of the imaging system, can be computed from object positions and from the optical system's parameters. Iterative deconvolution algorithms, which are necessary for volume reconstruction from microscopic image sequences, require considerable computation time. Using CNN-UM chips for these deconvolution tasks, a much higher speed, even real time processing seems to be achievable.