Mutual Information Criterion-Based Optimal Scale Selection for Image Denoising and Segmentation

Abstract

Scale-spaces play an important role in many computer vision tasks. Automatic scale selection is at the foundation of multiscale image analysis, but its performance is still very subjective and empirical. To automatically select the appropriate scale for a particular issue, a scale selection model based on information theory is proposed in this paper. The proposed model utilizes mutual information as a measuring criterion of similarity for the optimal scale selection in multi-scale analysis, with applications to image denoising and segmentation. Firstly, we focus on the morphological operator based scale selection to image denoising. This technique does not require the prior knowledge of the noise variance and can effectively eliminate the variation of illumination. Secondly, we develop a clustering based unsupervised image segmentation algorithm by recursively pruning the Huffman coding tree. The proposed clustering algorithm can preserve the maximum amount of information at a specific clustering number from the information-theoretical point of view. Finally, for the feasibility of the proposed algorithms, its theoretical properties are analyzed mathematically and its performance is tested by a series of experiments, which demonstrate that it yields the optimal scale for our developed image denoising and segmentation algorithms.