Сегментація зображень за допомогою нейромережевого підходу

Abstract

In this master's qualification work the analysis of image segmentation methods was carried out, the review of convolutional neural networks for the purpose of image segmentation was performed, the software tools for image segmentation using deep neural networks were substantiated and selected, and the study of the semantic segmentation of images using the AlexNet-FCN model on educational Picture Sets from the PASCAL VOC and NVIDIA-Aerial Drone Dataset. For research in the work used the environment of training deep neural networks DIGITS, as a library of learning chosen library Caffe. As a result of the training, the level of image segmentation probability was 83.02% on the PASCAL VOC test set and 99.17% on the NVIDIA-Aerial Drone Dataset test suite. Also, research on the segmentation of real images with the use of trained models has been conducted.