文献类型：期刊文献

中文题名：基于Double-D算法的舌像检测

英文题名：Tongue image detection based on Double-D algorithm

作者：刘佳丽[1];孙自强[1]

机构：[1]华东理工大学化工过程先进控制和优化技术教育部重点实验室,上海200237

年份：2020

卷号：41

期号：7

起止页码：2025

中文期刊名：计算机工程与设计

外文期刊名：Computer Engineering and Design

收录：CSTPCD;;北大核心:【北大核心2017】；

基金：中央高校基本科研业务费专项基金项目(222201917006)。

语种：中文

中文关键词：密集连接;通道剪枝;特征融合;识别分类;中医舌诊

外文关键词：densely connection;channel pruning;feature fusion;identification and classification;traditional Chinese medicine tongue diagnosis

摘要：针对舌像纹理复杂细微的特点,以YOLOv3目标检测算法为基础,提出一种Double-D优化架构的改进算法。采用细粒度特征融合和多尺度预测,设计稠密直通映射连接优化模型底层图像的特征信息,更加精准实现小目标的识别分类;进一步根据稀疏权重进行通道剪枝,降低模型参数计算量,便于网络深化。将该算法应用到中医舌像的临床病例的识别过程,其结果表明,其在图像识别的准确性和泛化性能上相较现有方法有一定提升。
Aiming at the complex and fine texture of tongue image,an improved algorithm of Double-D optimization architecture was proposed based on YOLOv3 target detection algorithm.Fine-grained feature fusion and multi-scale prediction were adopted,and densely shortcut mapping was designed to connect the feature information of the underlying images of the optimization model,so as to realize the recognition and classification of small targets more accurately.Channel pruning was further preformed according to the sparse weights,which greatly reduced the calculation of model parameters and facilitated the deepening of the network.The algorithm was applied to the clinical case recognition process of tongue image for traditional Chinese medicine.The results show that it has certain improvement in the accuracy and generalization performance of image recognition compared with the existing methods.