摘要
目的:利用中药饮片图像中颜色的种类和分布特征,构建与尺度和旋转无关的颜色匹配模板,建立中药饮片的颜色表征及图像识别方法。方法:选取根茎类、花、种子和果实类中药饮片共20种,每种样品各选取相应的2个观察面摄取图像,经图像分割,RGB颜色模型转换为L~*a~*b~*等图像处理过程,提取各观察面的图像前景的颜色参数。2个观察面的颜色向量按降序排序,插值缩放至一定尺度,按1∶1的权重构造综合颜色向量。对于具有向心分布的观察面(如横切面)图像,采用腐蚀操作由外至内逐圈提取各环带的颜色分量,并进行排序、缩放操作。以综合颜色向量作初始模板进行训练,计算各样本与模板的相关系数,结合t检验对阳性样本进行区间估计,以总识别率为考察指标,确定最优模板尺度、环带宽度和训练量。结果:不同种类饮片训练后的综合颜色模板的可视化结果易于目视辨别;测试260个饮片样本,由a~*,b~*2个颜色分量构建的综合颜色模板的识别性能优于L~*,a~*,b~*3个分量的模板,其总识别率为95. 8%(249个/200个)。结论:整合中药饮片2个不同观察面的图像颜色特征以构建综合颜色特征向量,对于相同药用部位的样品和不同药用部位的样品均可获得较好的识别分类结果;该方法对样品的形状、取样部位及颜色的随机变化有较强的抗干扰能力。
Objective: To construct the color matching template irrelevant to size and rotation according to the types and distribution characteristics of colors in images of Chinese herbal pieces,in order to establish color characterization and image identification methods for Chinese herbal pieces. Method: Totally 20 types of Chinese herbal pieces were selected,including rhizomes,flowers,seeds and fruits. For each sample,two observation surfaces were selected to extract color parameters in foreground through image processing such as image segmentation,model transformation from RGB to L~*a~*b~*. Color vectors of the two observation surfaces were sequenced in a descending order,scaled to a certain size by interpolating,and combined into an integrated color vector in a weight ratio of 1 ∶ 1. As for centripetally distributed observation surface images( e. g. transverse section),corrosion operation was conducted to extract the color components of each ring from outer to inner by circles,which were then ordered and scaled. The integrated color vector was used as initial template for training,the correlation coefficient between each sample and the template was calculated,and the interval estimation of positive samples were carried out by t test. With the total recognition rate as an indicator,the optimal template dimensions,width of ring and training volume were ultimately determined. Result: The visualization results of the trained templates of the varied herbal pieces were easy to be visually distinguished. After 260 samples of the herbal pieces were tested, the template of a and b components was better than that of L~*,a~*and b~*in terms of recognition performance,with a~*recognition accuracy of 95. 8%( 249/200). Conclusion: Color characteristics of images from two observation surfaces of Chinese herbal pieces are integrated to obtain the combined color feature vector,so as to achieve preferable recognition results for samples from both the same and different medicinal parts.This method boasts a strong anti-interference ability of random variation of sample shape,sampling part and color.
作者
陈仕妍
卢文彪
王凤梅
CHEN Shi-yan;LU Wen-biao;WANG Feng-mei(School of Pharmaceutical Sciences,Guangzhou University of Chinese Medicine,Guangzhou 510006,China)
出处
《中国实验方剂学杂志》
CAS
CSCD
北大核心
2020年第6期158-162,共5页
Chinese Journal of Experimental Traditional Medical Formulae