Entropy Regularization, Automatic Model Selection, and Unsupervised Image Segmentation

Abstract

In region-based image retrieval, the key problem of unsupervised image segmentation is to automatically determine the number of regions for each image in a database. Though we can solve this kind of model selection problem with some statistical criteria such as the minimum description length (MDL) through implementing the EM algorithm, the process of evaluating these criteria may incur a large computational cost. From competitive learning perspective, some more efficient approaches such as rival penalized competitive learning (RPCL) have also been developed for unsupervised image segmentation. However, the segmentation results are not satisfactory and the object of interest may be merged with other regions, since the RPCL algorithm is sensitive to the rival learning rate. In order to solve such problems, we then propose an iterative entropy regularized likelihood (ERL) learning algorithm for unsupervised image segmentation based on the finite mixture model, which can make automatic model selection through introducing entropy regularization into maximum likelihood (ML) estimation. Some segmentation experiments on the Corel image database further demonstrate that the iterative ERL learning algorithm outperforms the MDL based EM (MDL-EM) algorithm and the RPCL algorithm, and leads to some promising results.