Automatic Scene Classification: Algorithms and Applications

by

Andreas Savakis, Ph.D.
Associate Professor and Department Head
Department of Computer Engineering
Rochester Institute of Technology

Automatic scene classification is an image understanding problem that has applications in image database indexing and retrieval, selective image enhancement and multimedia applications, such as automatic albuming.  In this talk, a computationally efficient approach to scene classification is presented, where low level and semantic features are combined for improved performance. The first classification stage is based on low level features and k-nearest neighbor or Support Vector Machine classifiers. Low level features include color histograms and wavelet texture features, while semantic features include sky and grass information about the image. Knowledge integration of low level semantic features is achieved via a Bayesian network during a second stage of processing.  Results are presented for a database of 1300 photographic images and the overall classification performance is over 90 percent. Future work aims at employing similar algorithms and high performance environments for other computationally intensive image and video understanding tasks.