Natural scene classification using overcomplete ICA

Principal component analysis (PCA) has been widely used to extract features for pattern recognition problems such as object recognition [Turk and Pentland, J. Cognitive Neurosci. 3(1) (1991)]. In natural scene classification, Oliva and Torralba presented such an algorithm in Oliva and Torralba [Int. J. Comput. Vision 42(3) (2001) 145-175] for representing images by their ''spatial envelope'' properties, including naturalness, openness, and roughness. Our implementation closely matched the original algorithm in accuracy for naturalness classification (or ''manmade-natural'' classification) on a similar (Corel) dataset [Dong and Luo, Towards holistic scene descriptors for semantic scene classification, Eastman Kodak Company Technical Report, October 1, 2003]. However, we found that consumer photos, which are far more unconstrained in content and imaging conditions, present a greater challenge for the algorithm (as they typically do for image understanding algorithms). In this paper, we present an alternative approach to more robust naturalness classification, using overcomplete independent components analysis (ICA) directly on the Fourier-transformed image to derive sparse representations as more effective features for classification. Using both heuristic and support vector machine classifiers, we demonstrated that our ICA-based features are superior to the PCA-based features used in Oliva and Torrabla [Int. J. Comput. Vision 42(3) (2001) 145-175]; Dong and Luo [Towards holistic scene descriptors for semantic scene classification, Eastman Kodak Company Technical Report, October 1, 2003]. In addition, we augment ICA-based features with camera metadata related to image capture conditions to further improve the performance of our algorithm.