Analysis of Results of Scaling Digital Images by Interpolation Algorithms

Abstract

Scaling of digital bitmap images is often used in modern computer and telecommunications systems. Among image scaling algorithms, the most common are interpolation algorithms, namely nearest neighbor, Lanczos, bilinear and bicubic interpolations. However, in the process of scaling images by interpolation algorithms, characteristic distortions appear on them. Therefore, in this work, software implementation and research of image interpolation algorithms was performed in order to identify their advantages and disadvantages, areas of use and ways of improvement. In order to study the interpolation algorithms, a scaled fRGBs image was calculated based on the initial image fRGB, and then a scaled image fRGBs2 with the dimensions of the original image was calculated based on fRGBs. The scaling accuracy is evaluated using the Root Mean Square Error between the pixel values of the original and scaled images. The software for scaling images is developed in the Python language. Calculation of scaled images is performed by the cv2.resize() function of the OpenCV library. Using the developed program, scaling of a series of 100 images was carried out, the accuracy and speed of digital image scaling by interpolation algorithms were investigated. For each interpolation method, the average scaling error AR and the limits of its confidence interval ARmin and ARmax with a specified reliability γ are calculated. The average scaling time AT of a series of 100 images by different algorithms was determined. The research results showed that in most cases the smallest scaling error is provided by the bicubic interpolation algorithm, which is slightly inferior in speed to the nearest neighbor and bilinear interpolation algorithms. Recommendations for the application of interpolation algorithms have been developed. It is shown that the use of convolutional neural networks is promising for the highest quality image scaling.