中药材粉末显微特征图像的检测方法研究  

STUDY ON DETECTION METHOD FOR MICROSCOPIC IMAGES OF POWDERED CHINESE MEDICINAL MATERIALS

作　　者：王一丁[1] 孙霞[1] 李耀利 蔡少青[2] 袁媛[3] Wang Yiding;Sun Xia;Li Yaoli;Cai Shaoqing;Yuan Yuan(North China University of Technology,Beijing 100144,China;Peking University,Beijing 100041,China;National Resource Center for Chinese Materia Media,China Academy of Chinese Medical Sciences,Beijing 100700,China)

机构地区：[1]北方工业大学,北京100144 [2]北京大学,北京100041 [3]中国中医科学院中药资源中心,北京100700

出　　处：《计算机应用与软件》2022年第4期80-87,共8页Computer Applications and Software

基　　金：中医药行业科研专项(201407003);中央本级重大增减支项目(2060302)。

摘　　要：针对中药材粉末显微特征图像存在的目标断裂残缺这一关键问题,提出一种改进型SSD检测算法,即在SSD网络的预测卷积特征图之后加入SE模块,使SSD网络对该特征图的多个特征通道进行重要性的学习,并据此让包含较多信息且对最终结果起重要作用的特征通道分配到较大的权重。这样,断裂残缺的目标所保留的关键信息在网络迭代过程中能够被网络充分地学习,实现网络对目标的自动定位和种类识别,提高最终的检测效果。将改进型SSD算法用于厚壁细胞、导管和花粉孢子这三类显微图像的检测,mAP由79.8%提升到81.5%,这一结果证明了改进型SSD算法的有效性。The microscopic feature images of powdered Chinese medicinal materials have the problem of target fracture and incompleteness. For this key problem, this paper proposes an improved SSD detection algorithm. The SE module was added after the predictive convolution feature map of the SSD network, which made the network automatically learn the importance degree of each feature channel. And then, the feature channel which contained more information and played an important role in the final result got a greater weight, and made the key information retained by the fractured and incomplete target be fully learned by the network, which enabled the network to complete the automatic positioning and category identification of the target. So the accuracy of detection was finally improved. The improved SSD algorithm was applied to the detection of three types of microscopic images of thick-walled cells, ducts and pollen spores. The mAP increases from 79.8% to 81.5%. The experimental result demonstrates the effectiveness of the improved SSD detection algorithm.

关 键 词：中药材粉末 显微特征图像检测 SSD算法 SE模块 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]