基于改进YOLOv5的眼睛及瞳孔检测算法

Eye and Pupil Detection Algorithm Based on Improved YOLOv5

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。

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,and three image enhancement methods are compared.It is decided to use CLAHE(limited contrast Adaptive histogram equalization)method with good effect to enhance the image and improve the contrast;Secondly,the Swin Transformer module is introduced into YOLOv5 network to replace the last C3 module of the backbone network and three C3 modules in the three prediction heads,so as to improve the feature extraction ability of the network and improve the detection accuracy of eye parts;Finally,by introducing a multi-scale feature cross layer fusion mechanism in the YOLOv5 network,two target prediction heads are added to reduce the network's missed detection rate for eye and pupil regions.This article selected 2400 eye datasets with different levels of illumination from the Data setXVIII in the ELSE standard dataset,of which 1600 were training sets and 800 were testing sets.The experimental results show that the improved YOLOv5 network can detect the entire part of the eye and the complete pupil,with a high detection confidence.The mAP has increased by 3.2 percentage points,the Recall has increased by 2.7 percentage points,and has good real-time performance.