Image Comparison Based On Local Pixel Clustering

Image comparison is a fundamental step for monitoring images and has wide applications in many disciplines of sciences, including satellite imaging, medical research, quality control and so forth. This problem, however, is complicated because (i) the observed images often contain noise, (ii) the image intensity functions are discontinuous and have spatial structures. In the literature, a vast majority of the methods are intensity-based. However, such an approach is often questionable in real life situations where small changes in the background may not indicate an actual meaningful change in the images as long as the boundaries of the image objects remain the same. In this article, we propose a flexible and effective image comparison method based on local pixel clustering and construct a test statistic based on the <i>Variation of Information</i> metric. This is a feature based image comparison technique where edges or the jump points are considered as the primary features. Numerical examples and statistical properties show that the proposed image comparison method performs well in various real life scenarios.

图像对比是图像监测的基础性环节，在卫星成像、医学研究、质量控制等诸多科学学科中拥有广泛应用。然而该问题存在显著复杂性：其一，观测得到的图像通常含有噪声；其二，图像强度函数不连续且具备空间结构。现有研究中，绝大多数方法均基于图像强度展开。但在实际场景中，这类方法往往存在合理性争议：当图像目标的边界保持不变时，背景的微小变化并不代表图像存在具有实际意义的实质性变化。本文提出一种基于局部像素聚类的灵活高效的图像对比方法，并基于信息变异性（Variation of Information）指标构建检验统计量。该方法属于基于特征的图像对比技术，以边缘或跳变点作为核心特征。数值算例与统计特性分析表明，所提出的图像对比方法在各类实际场景中均表现优异。