Application of Convolutional Neural Network for Semantic Segmentation of Bandung Urban Scenes

Semantic segmentation is the process of classifying each pixel in an image into one of the predesignated classes. In its application for self-driving systems, semantic segmentation is used to identify objects and road conditions to make the right control decisions. This work uses the convolutional neural network method to perform semantic segmentation with the DeepLabV3+ architecture. Object of study is the urban scene dataset called Bandung Cityscapes collected around the city of Bandung, West Java, Indonesia and annotated by hand. Architecture testing is performed by combining the values of epoch, learning rate, and number of convolution filters to find the best mean intersection over union accuracy. The most optimal model achieved the highest average accuracy of 77.431% on the validation set, attained by the model with 128 convolutional filters, trained with 300 epochs and 0.0001 (10–⁴) learning rate. This model maintained an average training accuracy of 87.194%, average validation accuracy of 75.274%, and average difference between the two of 11.919% throughout the training process.