基于光谱与图像联合特征分析的色诊关键技术研究

Color inspection is one of important parts of traditional Chinese medicine diagnosis by inspection. At present, the color classification of diagnosis is mainly based on digital image and spectroscopic techniques. The convenience of digital image acquisition, visualization is strong, but the multi-dimensional feature vector makes relatively low classification performance than reflectance spectral characteristics in traditional Chinese medicine clinic color inspection. Therefore, based on the color digital image acquisition and diagnostic classification technique, establishing the inversion method from digital image to the reflectance spectral and joint features classification method of digital image and spectral reflectance become a key technology to improve the classification performance of traditional Chinese medicine color diagnosis. In this study, face and tongue color as a starting point, based on the “expert judgment, digital image color class-feature and reflectance spectra feature", it makes face and tongue color training set. Face and tongue digital image color features of are corrected by standard light environment. Transforming visible spectral reflectance features from the digital image features, and the establishment of color inspection classification method based on the visible spectral reflectance characteristics machine learning, it is forming joint features based on digital image and visible reflectance spectra of color classification. Research on joint features of digital image and spectral reflectance characteristics takes advantages of digital image information acquiring convenience and hyper spectral reflectance classification outstanding performance, in order to provide new ideas and methods for the objective study of TCM color inspection。

色诊技术是中医望诊现代化研究的重要内容。以往研究主要基于图像技术展开，操作便捷、可视化强，但存在条件依赖性及色阈范围局限性的不足，光谱技术可以为图像技术提供有效补偿，满足中医色诊高精度、高灵敏度需求。本研究以面色为切入点，筛选面色的Lab、HIS等色空间范围，基于光谱重建技术，通过标准光源校正的面色图像特征分析可见光谱反射率特征，研究肤色图像特征与光谱反射特征对应关系。采集临床典型面色的图像及光谱信息，建立“数字图像-反射光谱”特征联合约束关系，采用主成分分析(PCA)及人工神经网络(ANN)等机器学习方法，针对色诊光泽分类、面色五色分类，建立数字图像和可见反射光谱联合特征的色诊分类模型，提高中医色诊图像分类精度。应用已有TFDA-2型面色诊仪，校验图像与光谱联合特征的分类模型，实现基于联合特征的色诊图像技术补偿方案。本研究为色诊标准化提供关键技术支持，突破中医望诊应用瓶颈。

本研究使用TFDA-1面诊仪及海洋光学Flame光谱仪建立色诊“图像-光谱”特征联合采集关键共性前置基础，联合采集健康体检人群、疾病患者的庭、阙、明堂、颏等色部信息，分析健康常色、疾病赤面色、疾病黄面色、疾病白面色、疾病青黑面色的数字图像色度特征与对应的380-780nm范围内光谱反射率特征及光谱色度特征，在CIE Lab色度空间中建立图像色度特征参照系、光谱色度特征参照系，明确典型常色、病色光谱反射率对应的特征波长范围，联合图像、光谱特征优势，能更全面表征色诊信息；本研究采用多重线性回归、多项式回归、最小距离查找表、中心化查找表等明确明堂色部的图像色度、光谱色度对应关系，建立“图像-光谱”色度映射模型和“图像-光谱”色度查找表，分别将图像色度与光谱色度的色差减小至2.64±1.36、2.79±1.46、3.48±2.07、3.26±1.83，映射模型可补偿图像特征精度，促进色诊标准化；本研究采用主成分分析提取多维度图像特征、光谱特征的联合特征，根据典型面色光谱特征特点优化分类方案，采用人工神经网络建立“图像-光谱”联合特征色诊分类模型，将常色、病色分类准确率提高至86.64%，将疾病白面色、疾病青黑面色、疾病黄赤面色分类准确率提高至80.76%，将疾病黄面色、疾病赤面色分类准确率提高至88.38%，提升色诊分类准确率。本研究为色诊标准化提供关键技术支持，突破色诊精确表征关键技术瓶颈，提供图像、光谱特征联合表征中医典型面色的客观参考和科学依据。