[1]张 杨,辛国江,王 鑫,等.基于改进的 YOLOv5 网络的舌象检测算法[J].计算机技术与发展,2024,34(02):156-162.[doi:10. 3969 / j. issn. 1673-629X. 2024. 02. 023]
　ZHANG Yang,XIN Guo-jiang,WANG Xin,et al.Tongue Image Detection Algorithm Based on Improved YOLOv5 Network[J].,2024,34(02):156-162.[doi:10. 3969 / j. issn. 1673-629X. 2024. 02. 023]

点击复制

基于改进的 YOLOv5 网络的舌象检测算法()

分享到：

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

34

期数:

2024年02期

页码:

156-162

栏目:

人工智能

出版日期:

2024-02-10

文章信息/Info

Title:

Tongue Image Detection Algorithm Based on Improved YOLOv5 Network

文章编号:

1673-629X(2024)02-0156-07

作者:

张 杨; 辛国江; 王 鑫; 朱 磊

湖南中医药大学 信息与工程学院,湖南 长沙 410208

Author(s):

ZHANG Yang; XIN Guo-jiang; WANG Xin; ZHU Lei

chool of Information and Engineering,Hunan University of Chinese Medicine,Changsha 410208,China

关键词:

舌象检测; YOLOv5; ReLu 激活函数; 轻量化; SimAm 注意力机制

Keywords:

tongue image detection; YOLOv5; ReLu activation function; lightweight; SimAm attention mechanism

分类号:

TP391

DOI:

10. 3969 / j. issn. 1673-629X. 2024. 02. 023

摘要:

针对目前舌象检测模型在自然状态下对舌象检测存在的误检和漏检的问题,以收集的舌象为研究对象,提出了一种基于 YOLOv5 的自然状态下的舌象检测算法。 首先,将原有的 SiLU 激活函数替换为 ReLu 激活函数,减少指数运算,加速舌象检测网络收敛;然后,利用 Ghost 轻量化模块技术,大幅降低舌象检测网络的参数量;最后,将 SimAm 注意力机制融入特征提取网络获取舌象特征,从多维度融合舌象特征,降低自然环境对舌象特征提取的影响。得到一个轻量化的舌象检测模型,在自制的数据集上分析可知:轻量化检测模型参数量达到 7. 8 MB,检测的精度达到 96. 6% ,同时每秒处理帧数高达 86 帧,更适合自然状态下舌象的采集工作。 实验结果表明,改进的舌象检测网络在自制舌象数据集上,相比于其它常用检测算法,性能指标上均有不同程度提升,对舌象的检测效果更好。

Abstract:

Aiming at the problem of false detection and missed detection of tongue image in the natural state of the current tongue imagedetection model,we propose a tongue image detection algorithm based on YOLOv5 in the natural state with the collected tongue image asthe research object. Firstly,the original SiLU activation function is replaced with the ReLu activation function to reduce the exponentialoperation and accelerate the convergence of the tongue image detection network. Then,Ghost lightweight module technology is used togreatly reduce the number of parameters of the tongue detection network. Finally,the SimAm attention mechanism is integrated into thefeature extraction network to obtain tongue features,and the tongue features are fused from multiple dimensions to reduce the influence ofthe natural environment on the extraction of tongue features. A lightweight tongue image detection model?

is obtained, which can beanalyzed on the self-made dataset:the weight of the lightweight detection model reaches 7. 8 MB,the detection accuracy reaches 96. 6% ,and the number of frames per second is as high as 86 frames,which is more suitable for the collection of tongue images in the naturalstate. The experimental results show that compared with other commonly used detection algorithms, the performance index of theimproved tongue image detection network has been improved to different degrees,and the detection effect of tongue image is better.

常用功能