偏振测量作为一种无损伤、非标记、无接触、易扩展的光学技术，能够获取物体的各向异性结构信息，在海洋藻类观测方面有着独特的应用优势。穆勒矩阵包含了测量目标的所有偏振信息，是偏振光学领域的重要描述方法和研究对象。穆勒矩阵显微成像技术在研究微米尺度海洋藻类的各向异性特征问题上具有较大的应用潜力。本论文的工作以穆勒矩阵显微成像技术为基础，利用双波片旋转法与空气校准法实现海洋藻类穆勒矩阵的测量与校准，对数量达到万级的海洋藻类样本进行穆勒矩阵显微成像，建立了相应的样本处理方案和数据处理流程，对多种海洋藻类样例的偏振效应进行了分析。所测样本主要来自深圳周边海域，包括多种赤潮藻与非赤潮藻，在微观尺度上具有丰富的形态结构，是海洋生态研究中具有较大代表性和研究意义的海洋藻类。基于深度学习等多种算法，本论文着重研究海洋藻类穆勒矩阵的特征提取与分类问题，充分考虑了高维偏振数据处理的技术难点以及海洋藻类穆勒矩阵显微成像数据的特点，以较低的模型复杂度设计了多种适用于海洋藻类穆勒矩阵的深度卷积神经网络架构，在分类正确率与F1分数等评价指标上取得了具有实用意义的分类效果，演示了偏振测量技术在海洋藻类分类问题中的有效性。在海洋藻类穆勒矩阵偏振特征提取问题中，本论文从统计观点出发，以统计矩对海洋藻类穆勒矩阵的统计分布信息进行定量描述并初步评估问题难度，对比多种度量学习算法的技术路径和偏振特征提取效果，引入孪生网络算法并以分类问题的评估方式和指标对算法效果作定量比较，通过反复测试海洋藻类样本数据和多种孪生网络架构验证了孪生网络算法适用于处理海洋藻类偏振特征提取问题，以Pearson相关系数分析孪生网络首层特征与穆勒矩阵各阵元的统计关联，总结了相关实验现象和技术经验。本论文将穆勒矩阵成像技术与深度学习算法进行了有效结合，基于卷积神经网络与孪生网络等深度学习算法建立了一套适用于海洋藻类偏振特征提取与分类的大数据方法，充分利用偏振测量技术所提供的信息辅助研究包含复杂各向异性生物体系的海洋生态学课题。
As a damage-free, label-free, non-contact and easily expandable optical technology, polarimetric technology is capable of capturing anisotropic structural information of objects, therefore it has unique application adavantages in observing marine algae. The Mueller matrix, which is an important descriptor and research target in the field of polarization, contains all the polarimetric information of the mearsured object. The Mueller matrix microscopic imaging technology reveals great application potentials in studying the anisotropic characteristics of micro-sclae marine algae.Based on the Mueller matrix microscopic imaging technology, this thesis implements the measurement and calibration of the marine algal Mueller matrices by the double wave-plate rotating method and the correspondding air calibration. More than 10,000 marine algal samples have been mearsured by the Mueller matrix microscope, and the corresponding sample processing scheme and data processing pipeline are established. The polarimetric effects of various classes of marine algal samples are analyzed. The measured marine algal samples, which have study significations in marine ecological researches, are mainly collected in the surrounding waters of Shenzhen, including several classes of red tide algae and non-red tide algae that vary morphologically in microscale. With the help of various algorithms such as deep learning, this thesis focuses on the feature extraction and classification of marine algal Mueller matrix, and fully considers the technical difficulties of high-dimensional polarimetric data processing and the characteristics of marine algal Mueller matrices. A variety of deep convolutional neural network architectures are designed with low model complexity, and practical classification results are obtained on the classification accuracies and F1 scores.The effectiveness of polarimetric technology in the classification of marine algae has been demonstrated.In the problem of polarimetric feature extraction of marine algal Mueller matrix, this thesis starts from the view point of statistics, and quantitatively describes the statistical distribution information of marine algal Mueller matrix with statistical moments as a preliminary assessment of the difficulty of the problem, then compares the technical path and the algorithmic results of polarimetric feature extraction of multiple metric learning algorithms. The siamese network algorithm is introduced, and the algorithmic results are compared quantitatively with the evaluation methods and indicators in the framework of classification problems. By repeatedly testing the polarimetric data of marine algal samples and various siamese network architectures, the siamese network algorithm proves to be suitable for the problem of polarimetric feature extraction of marine algae. The Pearson correlation coefficient is introduced to analyze the statistical correlation between the features in the first layer of the siamese network and the Mueller matrix elements, and the relevant experimental phenomena and technical experience are summarized.In this thesis, the Muller matrix imaging technology and the deep learning algorithm are combined effectively. Based on the deep learning algorithms such as the convolutional neural networks and the siamese networks, a big data method that is suitable for the polarimetric feature extraction and classification of marine algae has been established. With the information provided by polarimetric technology, this big data approach may help study the researches in marine ecology including complex anisotropic biological systems.
