10ARSPC Template for Proceedings on CD ROM

Automated building recognition remains an open research problem in digital photogrammetry. Despite considerable research in this area over the past 15 years, no general methods have been found. Over the same period, considerable advances have occurred in aspects of computer vision that appear to have similar characteristics to the problems in photogrammetry. One approach to object recognition that has not been extensively explored by the photogrammetric community is machine learning. These techniques are based on principles similar to supervised classification, where the computer system is ‘trained’ to recognize certain features in the image space. While these techniques cannot provide all the answers, they have been used successfully in other vision problems to mimic the early stages of visual processing. Most computer vision problems require salient features to be detected in the image as part of the processing strategy. In this application, a simple form of wavelet analysis is used. The images are characterized using the wavelet analysis to provide multiresolution data for the machine learning phase. The learning phase is a binary classification problem, where the program is trained to recognize building and nonbuilding images. The learning in this situation is complicated by the high dimensionality of the image data and its wavelet derivatives. While many different classifiers could be used, Support Vector Machines (SVMs) are a relatively new classification tool that appear well suited to this application. SVMs take advantage of high-dimensional feature space to separate data that might be otherwise inseparable. This paper describes the use wavelet analysis and support vector machines to recognize buildings in aerial photographic images and reports the results of several evaluations that were undertaken.