基于互信息准则的图像平滑和分割

Image Denoising and Segmentation Based on Mutual Information Criterion

在计算机视觉领域,尺度空间扮演着一个很重要的角色。多尺度图像分析的基础是自动尺度选择,但它的性能非常主观和依赖于经验。基于互信息的度量准则,文章提出了一种自动选取最优尺度的模型。首先,研究专注于基于形态学算子的多尺度图像平滑去噪方法,这种技术不需要噪声方差的先验知识,可以有效地消除照度的变化。其次,通过递归修剪Huffman编码树,设计了一个基于聚类的无监督图像分割算法。一个特定的聚类数从信息理论的角度来看,提出的聚类算法可以保留最大的信息量。最后,用一系列的实验对算法的性能进行了验证,并从数学上进行了详细的证明和分析,实验结果表明本文提出的算法能获得最优尺度的图像平滑和分割性能。

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