Enhancing Semantic Segmentation Using Deep Neural Networks

Abstract

Deep neural networks (deep learning) are considered one of the most important methods used in solving computer vision problems.

Semantic segmentation algorithms are among the advanced algorithms in computer vision that provide high performance by giving them accurate details when used appropriate.

(U_Net) is a well-performing method in semantic segmentation problems, but it suffers from some problems due to its inability to deal with the instance

segmentation problems.

In this study, a modification and integration of a set of steps have been done in order to give U_Net the ability to deal with (instance segmentation), where two operations were added when pre-processing the data, namely cutting images into slices and encoding them according to the categories in them, in addition to segmenting images into Small equal parts to be able to train the model later

 (patchifying).

Also a special cost function when training the model (soft dice-loss) has been used and compared its performance with the most commonly used cost functions (cross-entropy and mae). The mentioned modifications were made to samples of the aerial images of specific areas and the ability of the modified model to identify the types of geographical areas included in the aerial image were tested, which were divided into five categories.