Contour-based feature extraction for image classification and retrieval

We propose a feature extraction scheme for application on image classification and retrieval that is based on shapes' contours, while discarding information within the boundaries such as colour and texture. The center of mass and opposite distances are calculated for every contour pixel and used to measure distances between pairs of images that are invariant to common transformations. We apply the k-nearest neighbours (k-NN) algorithm to classify/retrieve a query image according to the k closest images' classes. The resulting success rates were computed for the Kimia, MPEG-7 and Tari image data sets and compared with those of other techniques.