SINGLE IMAGE UPSCALLING METHOD USING MULTI-OUTPUT REGRESSION AND MOMENT METRIC

Abstract

The article proposes a new method for increasing the resolution of a single image, which solves the problem of image enlargement as a regression problem with several outputs. This method uses momentary metrics to evaluate the quality of magnified images. The proposed algorithm is implemented using a simple neural network with three hidden layers, and the results of numerical experiments on reference images have shown that image enlargement by this algorithm has a quality that exceeds the quality of enlarged images obtained by classical methods and the EDSR neural network. The proposed method is a promising solution for achieving super-resolution of a single image because it treats the problem as a machine learning problem, which provides greater flexibility and adaptability compared to traditional methods. In addition, the use of instantaneous metrics allows for a more accurate assessment of the quality of enlarged images. However, it should be noted that this study is limited by the number of reference images used in the experiments and the specific architecture of the neural network. In future research, it would be useful to investigate the performance of the proposed method on a larger dataset and to study different neural network architectures.