Mimicking human texture classification

In an attempt to mimic human (colorful) texture classification by a clustering algorithm three lines of research have been encountered, in which as test set 180 texture images (both their color and gray-scale equivalent) were drawn from the OuTex and VisTex databases. First, a k-means algorithm was applied with three feature vectors, based on color/gray values, four texture features, and their combination. Second, 18 participants clustered the images using a newly developed card sorting program. The mutual agreement between the participants was 57% and 56% and between the algorithm and the participants it was 47% and 45%, for respectively color and gray-scale texture images. Third, in a benchmark, 30 participants judged the algorithms’ clusters with gray-scale textures as more homogeneous then those with colored textures. However, a high interpersonal variability was present for both the color and the gray-scale clusters. So, despite the promising results, it is questionable whether average human texture classification can be mimicked (if it exists at all).