National Technical University of Athens (NTUA)
Διερεύνηση τεχνικών κατάτμησης εικόνας, υπολογιστικής νοημοσύνης και εμπείρων συστημάτων στην τηλεπισκόπηση
Abstract
dc:descriptionThe objective of this research was to research and implement state-of-the-art computer vision and machine learning methods for Object-Based Image Analysis (OBIA), as well as integration with knowledge-based expert systems. The first contribution was the development of a generic image segmentation algorithm, as a low level processing part of an integrated object-oriented image analysis system. The implemented algorithm is called MSEG and can be described as a region merging procedure. The second contribution of this research involved the design and development of a region-based multi-scale segmentation algorithm with the integration of complex texture features. The third contribution of this research involved the integration of Support Vector Machines (SVM) with OBIA. The Support Vector Machine is a theoretically superior machine learning methodology with great results in pattern recognition. Especially for supervised classification of high-dimensional datasets and has been found competitive with the best machine learning algorithms. In the past, SVMs were tested and evaluated only as pixel-based image classifiers. In the forth contribution of this research, an object-oriented image classification framework was developed which incorporates nonlinear scale space filtering into the multi-scale segmentation and classification procedures. Morphological levelings, which possess a number of desired spatial and spectral properties, were associated with anisotropically diffused markers towards the construction of nonlinear scale spaces. The fifth contribution of this research involved the development of a multiscale object-oriented image analysis framework, which incorporated a region merging segmentation algorithm enchanced by advanced edge features and nonlinear scale space filtering. The sixth contribution of this research involved the implementation of an object-based image classification method, incorporating the Relevance Vector Machine framework for image object classification. The Relevance Vector Machine (RVM) is a kernel classification method, extending the Support Vector Machine (SVM) through Bayessian theory, dealing with uncertainty within the kernel-based classification framework. Finally, in the seventh contribution of this research, a multimodal object-based image classification approach was developed and evaluated. The goal of this research was to integrate machine learning classification with knowledge-based expert systems, to extend the Object-Based Image Analysis methodology and to evaluate its effectiveness and prospects. The previously developed framework, based on scale-space representation and a state-of-the-art multi-scale image segmentation algorithm provided the primitive image objects, while their spatial and spectral properties formed a multidimensional feature space for the following classification steps. Then, a sparse kernel-based supervised classification method was developed, based on Relevance Vector Machines, which provided the initial spectral classification of the image primitives. This sparse classifier was able to perform with high accuracy even with a small number of sample objects while maintaining the accuracy performance of SVM methods. The classified image objects and their spatial and spectral features were represented and stored in a spatial database, using a vector representation in order to perform spatial analysis and reduce the memory footprint of the system. Then, a knowledge-based expert system was integrated into the spatial database to provide rule-based classification support through a pure expert system tool: CLIPS. This approach was evaluated against other object-based classification methods and some very promising experimental results were demonstrated. Overall, this disertation contributed in all parts of the Object-Based Image Analysis methodology, introducing and implementing state-of-the-art computer vision, image segmentation, machine learning and knowledge-based methods in the OBIA methodology, thus integrating new tools towards automating the classification and feature extraction tasks in remote sensing. This need for more accurate and automated classification is crucial in remote sensing today, since photointerpretation of big earth observation data is a tedious process. Future prospects of this research include the implementation of the proposed methods in various remote sensing applications.
Degree
thesis:*- Grantor dc:publisher
- National Technical University of Athens (NTUA)
- Year dc:date
- 2014
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Tzotsos, Angelos
- Τζώτσος, Άγγελος
Subjects
dc:subject × 20- Τηλεπισκόπηση
- Άντικειμενοστρεφής ανάλυση εικόνας
- Έμπειρα συστήματα
- Τεχνητή νοημοσύνη
- Κατάτμηση εικόνας
- Υπολογιστική νοημοσύνη
- Remote sensing
- Object-based image analysis
- Knowledge-based expert systems
- Artificial intelligence
- Image segmentation
- Machine learning
- Φυσικές Επιστήμες
- Γεωεπιστήμες και Επιστήμες Περιβάλλοντος
- Επιστήμες Μηχανικού και Τεχνολογία
- Άλλες Επιστήμες Μηχανικού και Τεχνολογίες
- Natural Sciences
- Earth and Related Environmental Sciences
- Engineering and Technology
- Other Engineering and Technologies
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/37425