Optimal Wavelength Range Selection by a Genetic Algorithm for Discrimination Purposes in Spectroscopic Infrared Imaging

When spectroscopic infrared imaging is applied to discriminate between different materials, multiple images have to be measured at different wavelengths or wavelength ranges. The time-consuming step in present on-line spectroscopic imaging is the measurement and processing time per identification of a number of spectroscopic images. If this number of images can be kept small, whereby an optimal discrimination is still guaranteed, the acquisition and processing time will be faster and, therefore, this approach becomes attractive in real-world applications. This paper describes the search for a limited number of spectroscopic wavelengths or wavelength ranges for images where optimal discrimination between the materials is guaranteed. This optimization is applied in particular to the discrimination between plastics and nonplastics. Because the number of potential wavelength combinations is huge, a genetic algorithm (GA) is used as a subset selection technique to solve this large-scale optimization problem. Since the problem concerns classification, a specific optimization criterion is developed. Finally, infrared images are measured at the calculated optimal wavelength ranges, and the resulting discrimination performance is compared with that of images measured at wavelengths chosen on the basis of a priori spectroscopic knowledge.