The work is devoted the study of the models of convolutional neural networks SRCNN, FSRCNN, and SubPixel CNN, which solve the Super-Resolution task, for finding the optimal parameters during the training stage. To evaluate the results, the standard PNSR metric was used, BSD was taken as the dataset. During analyzing activation functions and optimizers from the PyTorch library, the best combinations, that give a higher quality metric, were identified. Using the weights of the three best combinations obtained, a comparative characterization of the images belonging to the classes of text, face, night, nature, and the generated image was given. The conclusion is made about the evaluation of a particular architecture to obtain an image of a certain class. The finishing results can be used in further studies for improving the productivity and accuracy of SuperResolution methods.

Image Super-Resolution, convolutional neural network, peak signal-to-noise ratio, activation function, optimizer, SRCNN, FSRCNN, SubPixel CNN, PyTorch

Strokova A.V.,  ,  orcid.org/0000-0002-6850-610X,  ITMO University