<title>Color space quantization for inspection of textured objects</title>

Abstract

This paper describes an approach for recognizing naturally textured objects using color images. Natural objects, such as finished wood, yield images that are inherently difficult to analyze because large variations in visual appearance are common. In the application of interest here, traditional texture- and color-based techniques yielded poor results in our early experiments. However, we found that classification accuracy improved dramatically when a nonuniform quantization of the color space was chosen adaptively, using a set of training images. Ultimately, we developed a novel method for selecting a nonuniform partition of the color space so that differences between object classes are accentuated. The resulting partition serves as the domain for histograms of models and of observed images, and an information-theoretic similarity measure is used to perform recognition. The motivation for this system is to achieve high recognition accuracy in an industrial setting. Laboratory tests have demonstrated a high level of accuracy for this technique, even though the objects of interest exhibit large variations of texture and color.