Comparative analysis of semantic segmentation of terrestrial images using DeepLabv3+

Abstract

Feature extraction from images by semantic segmentation method with the help of deep
learning algorithms is one of the methods in the sub-discipline called computer learning.
Basically, the process is to give loaded data to deep artificial neural networks and train the
artificial neural network with this data again and again until it creates correct predictions. In
this study, DeepLab v3+ algorithm and two deep learning architectures such as Resnet18 and
Resnet50 were chosen as backbone for feature extraction task from terrestrial images. The
application was carried out on MATLAB. CamVid and Cityscapes datasets were used as
datasets. Among the models applied, the one with the highest evaluation accuracy is, where
the backbone is Resnet50 with 93.53% on Camvid and 89.29% on Cityscapes. The best results
were applied to the images, which were taken for the study outside the data sets, and the
results were evaluated visually.