Automatic selection of optical filters for classification in hyperspectral images

The color of a material is one of the most frequently used feature in automated visual inspection systems. While it is sufficient for many "easy" tasks, more complex materials such as food-stuffs and minerals usually require more complex features. Spectral ""signatures"" in the near infrared or UV spectrum have proven useful, but hyperspectral imaging devices are still too costly and too slow for industrial application. Therefore, off-the-shelve cameras and optical filters are used to extract characteristic features from the spectra. While the visual inspection community has acknowledged the benefits of this method, relatively few works are concerned with automatic selection of suitable filters. In a novel approach, filter selection is generalized as feature selection problem. In contrast to existing methods, this method can be used to select the best out of a large given set of filters, e.g. from a catalogue. This meta-method is exemplified by application of feature selection methods based on linear discriminant analysis, information theory and boosting.