ПРОГНОЗУВАННЯ ЯКОСТІ ВІДНОВЛЕННЯ ЗОБРАЖЕНЬ ІЗ ЗАСТОСУВАННЯМ ТЕСТУРНИХ ДЕСКРИПТОРІВ

Context. The problem of filling missing image areas with realistic assumptions often arises in the processing of real scenes in computer vision and computer graphics. To inpaint the missing areas in an image, various approaches are applied such as diffusion models, self-attention mechanism, and generative adversarial networks. To restore the real scene images convolutional neural networks are used. Although convolutional neural networks recently achieved significant success in image inpainting, high efficiency is not always provided.Objective. The paper aims to reduce the time consumption in computer vision and computer graphics systems by accuracy prediction of image inpainting with convolutional neural networks.Method. The prediction of image inpainting accuracy can be done by an analysis of image statistics without the execution of inpainting itself. Then the time and computer resources on the image inpainting will not be consumed. We have used a peak signalto-noise ratio and a structural similarity index measure to evaluate an image inpainting accuracy.Results. It is shown that a prediction can perform well for a wide range of mask sizes and real-scene images collected in the Places2 database. As an example, we concentrated on a particular case of the LaMa network versions although the proposed method can be generalized to other convolutional neural networks as well.Conclusions. The results obtained by the proposed method show that this type of prediction can be performed with satisfactory accuracy if the dependencies of the SSIM or PSNR versus image homogeneity are used. It should be noted that the structural similarity of the original and inpainted images is better predicted than the error between the corresponding pixels in the original and inpainted images. To further reduce the prediction error, it is possible to apply the regression on several input parameters