An MDL Approach to Color Image Segmentation

Abstract

In this paper, we propose an unsupervised colorimage segmentation method based on image features encoding. In this method, image segmentation is treated as minimizing the feature coding length. Image features are firstly obtained by any proper transformation that maps image data to feature space. Then a two-part MDL(Minimum Description Length) coding algorithm is proposed to encode image features: the histogram on each image channel is used to estimate the probability density function of the image features and the coding lengths foreach image channel and each partition border are summed together to determine the total coding length. A parameter free algorithm is proposed to minimize the MDL the coding length. We demonstrate that this algorithm can achieve good performance in visual evaluations with images from Berkeley image database.