Pupil Localization Method And Its Application In Strabismus Detection

The eye is an important organ for human beings to obtain external information,and the health of the eye is closely related to everyone’s life.As a common eye disease in children,strabismus has attracted great attention from the society.The incidence rate of strabismus in the world is 4%.If it is not treated in time in the early stage,it is easy to cause vision problems such as amblyopia,diplopia or stereopsis loss,and it will also have adverse effects on patients’daily life,mental health and academic employment.In order to treat patients with strabismus,rapid and accurate detection is very important.With the development of computer technology,image acquisition technology,digital image processing and artificial intelligence,how to use digital equipment to measure the degree of strabismus objectively and accurately has attracted the attention of extensive researchers.At present,most of the inspection methods commonly used in clinical strabismus specialties have subjectivity and other shortcomings.（1）Corneal light reflex testing:Generally,doctors only use it to roughly judge whether there is strabismus and the severity of strabismus,and it cannot accurately obtain the degree of strabismus.Besides,this method cannot eliminate the influence of kappa angle.（2）Cover test:It can accurately judge the direction of strabismus,but cannot obtain accurate display degrees.（3）Synoptophore:It can accurately measure the degree of strabismus,but the whole process requires the operation of professional doctors and the cooperation of the patients.The operation is complicated and the machine is expensive.Aiming at the problems existing in clinical strabismus measurement,this paper designs and implements a strabismus measurement system based on cover test.The main research contents of the thesis are as follows:（1）In the strabismus measurement system,we use a common near-infrared camera module.And the collected data usually has a lot of noise under the condition of less constraints on the subjects.Aiming at such situations,an improved pupil localization method based on Unet model is proposed.We incorporated several attention modules into central bottleneck part or skip connection part of original Unet model,which increases the model’s ability to learn pupil region features and suppress noise features.In addition,this method adds an ellipse fitting error loss termL_efeto the basis of BCE when training the network.This regularization term can effectively avoid network overfitting and improve model performance.The algorithm is trained and tested on the LPW public dataset.The results show that both the attention mechanism andL_efecan improve the detection accuracy comparing with original Unet model.When they are acted on Unet at the same time,the accuracy of this method is the highest within the error range of 1-15 pixels.Within five pixels of error,the detection rate increased by1.42%.In order to verify the accuracy of the proposed Unet-based improved pupil localization method in the strabismus system,we collected 3618 human eye images with a near-infrared camera module for testing.Within 5 pixel errors,our model’s accuracy rate can reach 93.7%.（4）In the strabismus system,the position of the human eye needs to be detected before pupil positioning.During the epidemic,in order to accurately detect the human eye area without relying on face detection,we use the YOLOv5s model for training.To this end,we integrated the public dataset Gaze Capture and the dataset collected and labeled by ourselves for training.The final F1 value reached 0.980 under the condition of confidence of 0.686,and the m AP reached 0.985 when the IOU threshold was 0.5.In order to improve the accuracy of the strabismus measurement system designed in this paper,the Hirschberg ratio was measured for each subject,and the empirical mean value was not used.Through multiple measurements on 6 subjects,we found that the variance of each subject’s Hirschberg ratio did not exceed 1.0,and the average R^2 of the fitted line could reach 0.99.In addition,in order to accurately detect the key frame after the removement of cover,according to the change of pupil size with incident light,the system detects the reflective point and pupil at the same time in the first three frames.We consider the frame with largest pupil radius to be the keyframe.Predicted keyframes are compared with manually labeled keyframes.Within 1 frame error,the average detection rate of key frames is 83.0%;within 2 frame errors,the average detection rate is 85.7%.In order to verify the accuracy of the designed strabismus measurement system,the subjects took a professional strabismus measurement in an eye hospital.We take this as standard,and statistics show that for healthy people,within 0.5 degrees of error,the average detection rate is 63.0%;Within 1 degree of error,the average detection rate is 83.3%.For strabismus patients without nystagmus,the errors were all within 1.5°,and the average error was 0.875°.In addition,the system can automatically save the Hirschberg ratio,calibration point detection results and strabismus measurement results of the subjects,and it also allows users to manually adjust the reflective point threshold to achieve the desired effect.