Image segmentation using deep learning methods

Abstract

This paper discusses the image segmentation methods based on deep learning methods. The segmentation object dataset was formed by two sets of World View 3 satellite images: RGB images and a 16-channel multispectral images. For dataset, I develop image preprocessing algorithms based on CNN. As segmentation methods Convolutional Neural Network are used due to its possibility to process not only by their spectral differences, but also by their spatial attributes. Based on U-Net, DeepLab and FullConv architecture networks were developed for satellite image segmentation. Finally, Jacard indexes of 3 networks were compared. These results are primarily due to the classes unevenness. To increase accuracy, it is necessary to train separately the classes sets. Among the applications of CNN in satellite images segmentation, we can distinguish urban infrastructure localization for Smart City technology, the segmentation of agricultural fields for the precision agriculture etc.