Research On Key Technology Of Hand-held Electronic Pupil Meter

The change of pupil diameter can not only reflect the change of external light intensity,but also reflect various states of the human body,especially those related to the nervous system.It is widely used in disease diagnosis,Sleep-deprived driving and other fields.The traditional method of pupil detection is through observation and calibration,usually using a caliper for manual measurement.Traditional pupil detection methods not only result in rough results,but also consume a lot of time and energy for doctors,reducing work efficiency and increasing the probability of errors.With the rapid development of embedded technology and computer vision algorithms,handheld electronic pupil meters have the characteristics of being small,accurate,and fast.They are increasingly widely used in medical,driving,and other fields,so fast and accurate pupil detection algorithms are very important.This article designs a two-stage pupil detection method,which first detects the human eye and then detects the pupil on top of the human eye.The main research content and achievements of this article are as follows:Firstly,a human eye detection method based on improved MTCNN was proposed,and in order to accelerate the detection speed,the landmark section was deleted;Optimized the input size of the network to make it more suitable for human eye detection;Replacing the pooling layer of the network with a convolutional layer to fully extract image features,experimental results show that compared to the original MTCNN algorithm,the improved MTCNN algorithm improves detection accuracy by about 3% at the expense of a small frame rate.The advantage of using this algorithm is its fast speed.Then,an improved algorithm YOLO-GG2 T based on YOLOv5 s was proposed.In order to accelerate the detection speed of the network,the Ghost module,Mobile Netv3 module,and Shuffle Netv2 module were used to improve the backbone part of YOLOv5 s.Overall,the Ghost module had the best performance;In order to further simplify the network,the target detection layer of the network has been changed from 3 to 2;In order to better extract features,Transfomer and Swin Transfomer were introduced into the original YOLOv5 s,respectively.Overall,the Transformer module performed better.The experimental results showed that compared to the YOLOv5 s model,YOLO-GG2T’s m AP＿0.5:0.95 increased by 0.3%,GFLOPs decreased by 52.5%,parameter file size decreased by 48.2%,and both accuracy and speed were effectively improved.The advantage of using this algorithm is its high accuracy.Finally,an improved algorithm U-Net-ERM based on U-Net was proposed.Firstly,the bottom of U-Net was optimized as a residual attention convolution module;Added ECA channel attention in the skip connection section,improving attention to pupil features;The decoding part was optimized as a strengthened feature fusion module,and the pupil features were fully obtained in the decoding part.The experimental results showed that on the CASIA Iris Distance dataset,compared to the U-Net model,the improved model increased m IOU by 0.57% and m Precision by1.78%.Use improved MTCNN for speed requirements and YOLOv5 s for accuracy requirements.