Automated Semantic Content Extraction from Images

• Main Author: Arab Khazaeli, Mahdi
• Other Authors: Knapp, Gerald
• Format: Others
• Language: en
• Published: LSU 2013
• Subjects: Engineering Science (Interdepartmental Program)
• Online Access: http://etd.lsu.edu/docs/available/etd-07052013-150623/

In this study, an automatic semantic segmentation and object recognition methodology is implemented which bridges the semantic gap between low level features of image content and high level conceptual meaning. Semantically understanding an image is essential in modeling autonomous robots, targeting customers in marketing or reverse engineering of building information modeling in the construction industry. To achieve an understanding of a room from a single image we proposed a new object recognition framework which has four major components: segmentation, scene detection, conceptual cueing and object recognition. The new segmentation methodology developed in this research extends Felzenswalb's cost function to include new surface index and depth features as well as color, texture and normal features to overcome issues of occlusion and shadowing commonly found in images. Adding depth allows capturing new features for object recognition stage to achieve high accuracy compared to the current state of the art. The goal was to develop an approach to capture and label perceptually important regions which often reflect global representation and understanding of the image. We developed a system by using contextual and common sense information for improving object recognition and scene detection, and fused the information from scene and objects to reduce the level of uncertainty. This study in addition to improving segmentation, scene detection and object recognition, can be used in applications that require physical parsing of the image into objects, surfaces and their relations. The applications include robotics, social networking, intelligence and anti-terrorism efforts, criminal investigations and security, marketing, and building information modeling in the construction industry. In this dissertation a structural framework (ontology) is developed that generates text descriptions based on understanding of objects, structures and the attributes of an image.