Automated classification of quilt photographs into crazy and non-crazy.

A. Gokhale and P. Bajcsy

Indian Institute Of Technology, Kharagpur (India) and NCSA, University of Illinois, Urbana, USA

IS&T/SPIE Electronic Imaging 2011
San Francisco, CA, January 23 - 27, 2011

This work addresses the problem of automatic classification and labeling of 19th- and 20th-century quilts from photographs, which are classified according to the quilt patterns into crazy and non - crazy categories. The motivation of our work is in automated annotation of a large collection of quilt images for research purposes of humanists. The research value of annotations for humanists is in understanding the distinct characteristics of an individual quilt-maker or relevant quilt-making groups in terms of their choices of pattern selection, color choices, layout, and original deviations from traditional patterns. The current annotation method is manual and the assignment is achieved by visual inspection. According to our knowledge, there does not exist currently a clear definition of the level of crazy-ness, nor an automated method for classifying patterns as crazy and non-crazy.

We approach the problem by modeling the level of crazy-ness by the distribution of clusters of color-homogeneous connected image segments of similar shapes. The model is turned into a set of image features that are extracted and represent our model of crazy-ness. The features are input into a supervised classification method, such as the Support Vector Machine (SVM) with the radial basis function kernel and optimized using 10-fold cross validation, used in our work. The classification methodology consists of four steps. In the first step, a color-homogeneous Region/Texel is detected by color based K-means clustering followed by connectivity analysis. This step leads to a cluster of color - homogeneous regions that represent quilt patches with similar colors. The second step uses a divide-and-conquer approach to identify sub-clusters that share similar shape properties such as area and perimeter. For each sub-cluster, statistics of nearest neighbor distances are computed for example, the mean and variance of distances. In the third step, a quilt signature per image is formed from parameters including the number of clusters containing a single color-homogeneous Region/Texel, number of clusters containing multiple color-homogeneous Region/Texel, the maximum and average number of color-homogeneous regions per cluster of regions and the minimum variance present in any cluster. Our selection of these parameters is based on the observation that crazy patterns have a small number of color-homogeneous and shape-similar regions in a cluster and a large number of clusters containing only a single region. They also have no symmetry and hence large variance in inter- Region nearest neighbor distance. In contrary, non-crazy patterns will have a small number of clusters and a large number of color homogeneous and shape-similar regions in a cluster. Finally, a Support Vector Machine (SVM) model is trained using labeled quilt images and 10-fold cross validation is used.

We implemented the classification methodology using a combination of Java and Matlab code. The algorithm was applied to 40 quilt images from the MATRIX database at the Michigan State University. We report almost 90 percent classification accuracy over 40 images using SVM and its radial basis function. In the future, we plan on extending the categorical model of crazyness to a continuous function reflecting the level of craziness.