Assessment of Feasibity of Neural Networks to Increase Image Clarity

The article describes the analysis of applying classical metrics for assessing the image quality with view of making a decision on a sucsessful solution of the SISR task. It presents the classification of the existing metrics for the assessment of images in general and of their clarity in particular. The majory of practical applications and app solutions in the present use generative and convolutional neural networks for solving the task of increasing the resolution of images. As a mechanism for enhancing the image resolution and clarity the neural netrworks of various architecture were implemented and used in the frame of the work: BSRGAN, SwinIR, ESRGAN, as well as the interpolation algorithm. The assessment of the images being restored was conducted through the subjective and objective analysises. The widely spread objective metrics for assesing the effectiveness of enhansing the resolution are the classical metrics PSNR and SSIM. It is shown that the application of the above metrics is not sufficiently accurate. The article cosideres such quality evaluation metrics as BRISQUE, NIQE, PIQUE, FDL and metrics that can be obtained based on the image similarity assesment. The results obtained make it possible to state that a really correct assessment of restored (enhanced) images should be conducted through a complex analysis with the use of several metrics of various focuse.