Eigenviews for object recognition in multispectral imaging systems

Abstract

We address the problem of representing multispectral images of objects using eigenviews for recognition purposes. Eigenviews have long been used for object recognition and pose estimation purposes in the grayscale and color image settings. The purpose of this paper is two-fold: firstly to extend the idealogies of eigenviews to multispectral images and secondly to propose the use of dimensionality reduction techniques other than those popularly used. Principal Component Analysis (PCA) and its various kernel-based flavors are popularly used to extract eigenviews. We propose the use of Independent Component Analysis (ICA) and Non-negative Matrix Factorization (NMF) as possible candidates for eigenview extraction. Multispectral images of a collection of 3D objects captured under different viewpoint locations are used to obtain representative views (eigenviews) that encode the information in these images. The idea is illustrated with a collection of eight synthetic objects imaged in both reflection and emission bands. A Nearest Neighbor classifier is used to perform the classification of an arbitrary view of an object. Classifier performance under additive white Gaussian noise is also tested. The results demonstrate that this system holds promise for use in object recognition under the multispectral imaging setting and also for novel dimensionality reduction techniques. The number of eigenviews needed by various techniques to obtain a given classifier accuracy is also calculated as a measure of the performance of the dimensionality reduction technique.