基于模式识别理论的相位不连续点检测及应用

Phase of light can be obtained by optical interferometry or phase retrieval from intensity images. In the process of obtaining phase, the phase discontinuity problem can not be ignored. No matter the 2π jumpers in wrapped phase maps or the real discontinuities of phase, identifying both kinds of phase discontinuities is a necessary step for further processing. Traditional phase discontinuity detection methods may get worse when data are noisy. Recently, researchers tried to use the neural network to identify the phase discontinuity, which got promising results. But this kind of classification-based method, which has not been studied thoroughly, needs tagged data to train the classifiers. This project will study the phase discontinuity problem under the framework of pattern recognition. The mainly research contents include three aspects. Firstly, for problem modeling of phase discontinuity detection, the clustering model, which doesn't need tagged data as training set, will be introduced as well as studying the classification model thoroughly. Secondly, the feature extraction on phase data will not only focus on distinguishing the phase discontinuous pixels from the continuous ones, but also recognition of the two kinds of phase discontinuities. Thirdly, for applications of phase unwrapping and discontinuous object measurement, specific algorithms which have a step of phase discontinuity detection are designed, and the effectiveness of these algorithms will be validated by shape measurement experiments via a Mach-Zehnder interferometer.

光的相位一般通过干涉方法间接测量或者利用相位恢复算法从强度图中恢复。在获得相位的过程中，相位不连续点是不容忽视的问题。无论是包裹相位中的2π相位跳变，还是待测相位本身的不连续性，准确检测出这两种不连续点是后续处理的必需环节。传统的相位不连续点检测方法易受噪声影响。近年来，有研究人员尝试用模式识别中的神经网络分类方法来检测相位不连续点，取得不错的结果，但这方面研究还不够深入，而且分类方法需要标注数据作为训练集。本项目将在模式识别框架下系统研究相位不连续点检测，研究内容包括：1)不连续点检测问题建模，深入研究分类模型的同时引入聚类模型，以应对标注数据很难获取的情况；2) 相位数据的特征提取方法，拟设计有针对的特征不仅表征出不连续性而且将两种不连续点区分开；3)结合相位解包裹和不连续表面测量两个应用，分别设计包含相位不连续点检测环节的算法，并通过基于马赫-曾德干涉仪的面形测量实验进行算法验证。

相位是光学测量中常借助的物理量。在获得相位的过程中，相位不连续点是不容忽视的问题。无论是包裹相位中的2π相位跳变，还是待测相位本身的不连续性，准确检测出这两种不连续点是后续处理的必需环节。本项目利用图像分析技术和模式识别技术对相位不连续点检测问题展开研究，主要研究内容和结果包括：1) 设计了一套基于聚类框架的相位不连续点检测方法，该方法利用一组图像边缘特征来描述相位不连续点，然后使用k-means算法将相位像素划分为“连续点”与“非连续点”两个集合，从而得到相位非连续点检测结果。该方法在模拟包裹相位图与模拟非连续镜面相位图上均取得良好的效果，分别检测出了两种相位不连续点。2) 基于包裹相位2π跳变点的数值特点，设计了一组专门针对包裹相位2π跳变点的图像滤波器，该组滤波器对包裹相位2π跳变点响应强烈，配合简单的阈值，即可较好地提取包裹相位2π跳变点，在模拟包裹相位图和真实移相干涉所得包裹相位图上效果好于基于图像边缘特征的聚类提取方法。3) 设计了一种基于相位不连续点检测结果的相位解包裹方法，在基于四步移相干涉的镜面测量结果的相位解包裹中取得良好效果。