Resumen
The purpose of this work is to study the effectiveness of augmentation methods of image sets when they are insufficient in training sample of neural networks for solving semantic segmentation problems. For this purpose, the main groups of augmentation methods were considered and their effectiveness in solving problems of semantic segmentation of medical images was investigated. Two deep architectures DeepLabV3+ with the EfficientNetB6 encoder were used for training, testing and validation. The Intersection over Union and Dice coefficient were chosen as the target metrics for comparing the quality of semantic segmentation of images, which made it possible to determine the models with the best predictions. The obtained results confirmed the effectiveness of the proposed set of augmentation methods. The result of the work was the creation of an effective approach to augmentation of medical image sets to solve the problem of semantic segmentation.