Building segmentation in remote sensing images using region merging approach with convolutional neural network-based model

Shoaib, Asim (2025) Building segmentation in remote sensing images using region merging approach with convolutional neural network-based model. Master dissertation/thesis, UTAR.

PDF Download (8Mb)

Abstract

Image segmentation is a process used to delineate objects in an image as regions of interest (ROIs). Poor delineation can lead to over segmentation (OS), resulting in creation of small regions that do not represent meaningful segmented ROIs. Region merging is one of the common approaches used to prevent OS in images. This approach iteratively merges adjacent regions based on merging criterion (MC) that defines the similarity of features between them. In the existing research works, feature map of labelled images were generated either manually or using a specialised software to derive MC. This process is labour intensive and time consuming. Therefore, in this research MC is derived with the assistance of convolutional neural network (CNN)-based deep learning model to perform region merging without any human intervention. In this research, AttentionU-Net model is used to generate feature map that is used to derive MC for merging building regions in WHU remote sensing images dataset. From experiments conducted, prominent features of building regions which are colour, texture, shape, and edges were extracted from the feature map viii to derive MC. This MC is used for merging the OS regions generated by simple linear iterative clustering (SLIC) algorithm. The proposed region merging approach has achieved an average F-measure of 0.91 in segmenting building regions in WHU remote sensing images. This is an improvement compared to previous research work on region merging, which achieved an average F-measure of 0.63 in delineating buildings regions in the same dataset. Moreover, the proposed region merging approach has achieved an average goodness of segmentation,

Actions (login required)