Algorithm Sib Image Unsupervised Classification

The IB principle is an extension of rate distortion theory. By defining a relevant variable Y with respect to source variable X, the IB method derives an appropriate distortion measure, which copes with the difficulty of choosing the distortion function in rate distortion theory. Along with the embedded applications of IB theory, some shortcomings of IB algorithms have emerged. One of these problems is that the relevant variable Y must appear in the form of co-occurrence of two variables X and Y. So we can't directly apply the IB algorithms to unsupervised image categorization where the images are in the raw pixel form.The sIB algorithm is one of the best and widely used IB algorithms. Aiming to apply the sIB algorithm to analyzing image data, this paper proposes a BoW-sIB algorithm for automatically categorizing images based on the image semantics. The BoW-sIB algorithm takes Bag-Of-Visual-Words model to represent each image, which extracts local features from each image and maps them into a visual vocabulary. Then, each local feature is labeled by the corresponding visual word, and the image data set is finally transformed into a co-occurrence matrix, a tally of the counts of each visual word in every image. The experimental results on Caltech data set show that the BoW-sIB algorithm can effectively reveal the original categories from the collection of unlabeled image data set and outperforms other unsupervised learning algorithms, including k-means, three latent variable models, two spectral clustering methods and Affinity Propagation algorithm.