IMAGE PREPROCESSING ALGORITHM TO REDUCE THE PROBABILITY OF OVERFITTING OF CONVOLUTIONAL NEURAL NETWORKS ON A NEURAL ACCELERATOR

The task of restoring lost fragments of monumental painting is relevant in the context of preservingcultural heritage sites. Modern artificial intelligence technologies, including convolutional neural networks(CNN), significantly expand the possibilities of restoration, allowing for the automation of compleximage restoration processes. In particular, the restoration of lost elements of frescoes requires preciseanalysis tools that can predict missing fragments with minimal errors, while preserving the artistic style ofthe original. The purpose of this study is to develop a technique of automated restoration of lost fragmentsof monumental painting images using CNN (using frescoes as an example). This goal was achieved bysolving the following problems: obtaining fresco images using appropriate methodological and technicaltools, applying the U-Net architecture for image segmentation and reconstruction, predicting lost areasbased on color characteristic analysis. The photogrammetry method and the designed device, which wereused to perform multi-angle shooting, provided high-quality source data for subsequent processing. Adaptationof the U-Net architecture to the image segmentation task has proven its effectiveness in identifyingkey structural elements of frescoes, which contributed to the accurate reconstruction of lost areas.To predict the lost areas, color characteristics were analyzed in the HSL system, which allowed the CNNto predict the missing colors with a high degree of accuracy. Brief conclusions of the study show that theproposed technique allows restoring both the shape and color of lost fragments of frescoes. The proposedtechnique is planned to be used for the restoration of other types of art works, which makes it promisingfor further research.