Biased clustering methods for image classification

Abstract

Classification of image data can be done using supervised or unsupervised methods. Each approach has advantages and disadvantages: supervised methods require labeled samples in order to create signatures or discriminating functions to classify unknown data, which at the end of the process will have class labels attached to it. Unsupervised methods, usually based on clustering, do not require samples for the classes, but their result will be unlabeled, requiring additional processing steps to attach labels to pixels on images. In this paper a new method for classification is presented, called biased clustering, which will use imprecise information about classes to create expectancies for assignment of a pixel to a class. These expectancies will be validated or corrected by a clustering method. The advantage over supervised methods is that the samples for the classes does not need to be precisely labeled, and can be derived with simple image processing methods. The advantage over basic clustering methods is that the pixels will be labeled at the end of the classification. An application of the method will be presented, and results will be discussed.