[1]韩慧妍*,范鑫茹.基于改进 YOLOv5 的眼睛及瞳孔检测算法[J].计算机技术与发展,2024,34(04):76-81.[doi:10. 3969 / j. issn. 1673-629X. 2024. 04. 012]
　HAN Hui-yan*,FAN Xin-ru.Eye and Pupil Detection Algorithm Based on Improved YOLOv5[J].,2024,34(04):76-81.[doi:10. 3969 / j. issn. 1673-629X. 2024. 04. 012]

点击复制

基于改进 YOLOv5 的眼睛及瞳孔检测算法()

分享到：

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

卷:

34

期数:

2024年04期

页码:

76-81

栏目:

媒体计算

出版日期:

2024-04-10

文章信息/Info

Title:

Eye and Pupil Detection Algorithm Based on Improved YOLOv5

文章编号:

1673-629X(2024)04-0076-06

作者:

韩慧妍¹; 2; 3* ; 范鑫茹¹; 2; 3

1. 中北大学 计算机科学与技术学院,山西 太原 030051;

2. 机器视觉与虚拟现实山西省重点实验室,山西 太原 030051;

3. 山西省视觉信息处理及智能机器人工程研究中心,山西 太原 030051

Author(s):

HAN Hui-yan1; 2; 3* ; FAN Xin-ru1; 2; 3

1. School of Data Science and Technology,North University of China,Taiyuan 030051,China;

2. Shanxi Key Laboratory of Machine Vision and Virtual Reality,Taiyuan 030051,China;

3. Shanxi Province’s Vision Information Processing and Intelligent Robot Engineering Research Center,Taiyuan 030051,China

关键词:

眼睛及瞳孔检测; YOLOv5; CLAHE; Swin Transformer; 多尺度特征跨层融合机制

Keywords:

eye part detection;  YOLOv5; CLAHE;  Swin Transformer;  Multi scale feature cross layer fusion mechanism

分类号:

TP391

DOI:

10. 3969 / j. issn. 1673-629X. 2024. 04. 012

摘要:

针对眼睛图像易受光照干扰导致的眼睛部位和瞳孔部位检测不准确及误检漏检的问题,提出基于改进 YOLOv5 的眼睛及瞳孔检测算法。 首先,进行图像预处理,对比了三种图像增强方法,决定运用效果较好的 CLAHE( 限制对比度自适应直方图均衡化) 方法进行图像增强,提高对比度;其次,在 YOLOv5 网络中引入 Swin Transformer 模块代替骨干网络的最后一个 C3 模块和三个预测头中的三个 C3 模块,提高网络的特征提取能力,提升眼睛部位的检测精度;最后,在 YOLOv5 网络中通过引入多尺度特征跨层融合机制的方法,增加两个目标预测头,降低网络对眼睛部位和瞳孔部位的漏检率。 该文从 ELSE 标准数据集中的 Data setXVIII 中选取了受光照程度不同的眼睛数据集 2 400 张,其中,1 600 张为训练集,800 张为测试集。 实验结果表明,改进后的 YOLOv5 网络能检测出眼睛整体部位及完整的瞳孔部位,检测置信度也较高,mAP 提高了 3. 2 百分点,Recall 提高了 2. 7 百分点,且具有较好的实时性。

Abstract:

To address the issue of inaccurate and missed eye and pupil detection caused by the susceptibility of eye images to light interference,an improved YOLOv5 based eye?

and pupil detection algorithm is proposed. First of all, image pre-processing is carried out, andthree image enhancement methods are compared. It is decided to use CLAHE?

( limited contrast Adaptive histogram equalization) methodwith good effect to enhance the image and improve the contrast; Secondly, the Swin Transformer module is introduced into YOLOv5network to replace the last C3 module of the backbone network and three C3 modules in the three prediction heads, so as to improve thefeature extraction ability?

of the network and improve the detection accuracy of eye parts; Finally, by introducing a multi-scale featurecross layer fusion mechanism in the YOLOv5 network, two target prediction heads are added to reduce the network’s missed detectionrate for eye and pupil regions. This article selected 2 400 eye datasets with different levels of illumination from the Data setXVIII in theELSE standard dataset, of which 1 600 were training sets and 800 were testing sets. The experimental results show that the improvedYOLOv5 network can detect the entire part of the eye and the complete pupil, with a high detection confidence. The mAP has increasedby 3. 2 percentage points, the Recall has increased by 2. 7 percentage points, and has good real-time performance.

常用功能