The Use of the Integral Method to Compare the Efficiency of Neural Networks to Improve Image Resolution Depending on Size and Format Using the Example of Road Photos

Roads occupy one of the most important places in the life of a modern person. The quality of the road surface is its key characteristic. To assess the quality of roads, there are many different systems that mainly solve the problem using image analysis. For better work with photos, they are preprocessed. Preprocessing, for example, includes changing the image resolution, which can be implemented using machine learning methods, including the use of neural networks. In this paper, a comparative analysis of the neural networks efficiencies to improve image resolution, depending on size and format, using the integral method is carried out. The neural networks selected are EDSR, ESPCN, ESRGAN, FSRCNN, LapSRN. The research material consists of road surface images in 5 sizes <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(\mathbf{100}\times \mathbf{100}, \mathbf{141}\times \mathbf{141},\ \mathbf{173}\times \mathbf{173},\ \mathbf{200}\times \mathbf{200},\ \mathbf{224}\times \mathbf{224})$</tex> and 3 formats (bmp, jpg, png). All neural networks have a linear O(n) dependence of the average time to obtain an improved image version on the photo size for each format. EDSR, ESPCN, FSRCNN, LapSRN also have a linear O(n) dependence for the average allocated memory on the size for all formats. ESRGAN has a constant O(1) dependence on the allocated memory for jpg and png formats, and almost constant dependence for bmp. ESPCN has the best time efficiency, and ESRGAN has the best space efficiency. ESRGAN and ESPCN also have the best integral efficiency values.