Region Segmentation and Annotation with Vehicle Detection Validation Application in Airborne Images

Abstract

In this work, the authors propose an automatic image segmentation and annotation system for airborne images. Initial region segmentation using existing region segmentation methods is applied to airborne images first. To deal with over-segmentation on the initial region segmentation results, the authors performed graph-based region merging by constructing an undirected-graph based on 8-connected local neighborhood. For each region, the authors extracted low-level features and used the Support Vector Machine (SVM) classifier to annotate the region with labels. Based on the output of the SVM classifier, adjacent regions with the same label were further merged to obtain the final segmentation and labeling result. The segmentation and annotation results are useful for vehicle detection validation. The experiments show that the proposed system can effectively segment and label various aerial images on a highly challenging dataset. Also, vehicle detection can be substantially enhanced with the help of the proposed annotation results and validation scheme.