DefinePK

Abstract

 We present a new method for image feature-extraction for learning image classification. An image is represented by a feature vector of distances that measure the dissimilarity between regions of the image and a set of fixed image prototypes. The method uses a text-based representation of images where the texture of an image corresponds to patterns of symbols in the text string. The distance between two images is based on the LZ-complexity of their corresponding strings. Given a set of input images, the algorithm produces cases that can be used by any supervised or unsupervised learning algorithm to learn image classification or clustering. A main advantage in this approach is the lack of need for any image processing or image analysis. A non-expert user can define the image-features by selecting a few small images that serve as prototypes for each class category. The algorithm is designed to run on a parallel processing platform. Results on the classification accuracy and processing speed are reported for several image classification problems including aerial imaging.