Aided image understanding system

Abstract

Tools for automatic image understanding for damage assessment for environmental catastrophes or military operations are essential for managing operator workloads. The paper proposes a tool which integrates image segmentation and classification with the goal of providing accurate and timely information about the areas of study. Traditional methods involving image segmentation followed by classification have not lived up to their potential due to the inherent semantic gap between these two functions. Segmentation algorithms have been limited in their success in extracting objects of interest which in turn limits classification performance since the segmentation algorithm has no a priori knowledge of the objects in the image. Segmentation algorithms fail in one of two directions: (i) over-segmentation where the object of interest is divided into many smaller regions or (ii) under-segmentation where the object of interest is merged with irrelevant background information. Both problems can confound the classification process. The approach is demonstrated on aerial images from the Katrina disaster to be able to detect buildings that may have been damaged or displaced from their original positions.