Research On Dry Eye Detection Algorithm Based On Deep Learning

Dry eye is a very common eye disease in ophthalmology clinics.The domestic invasive dry eye detection results in aggravation of the patient ’ s condition by dripping chemical reagents into the patient’s eyes.The foreign non-invasive dry eye detection uses traditional image processing methods to extract features on the basis of optical imaging.However,this method is slow and has low accuracy.Based on the above background,this article uses deep learning methods for dry eye detection.It is looked forward to automatically find the location of tear break up,and then compute the tear break up time to diagnose whether the patient has dry eye.This paper takes eye pictures based on placidio disk imaging as the research object to study and explore the dry eye detection model based on deep learning.The main research details are as follows:First of all,in conjunction with the Liaoning He Eye Industry Group,a large number of images were collected from the clinic,and a tear up image image database was constructed to solve the problem of lack of data in the tear break up image database.The tear break up position is marked with high quality;image segmentation,image enhancement,and remove noise are performed on the images in the image library.Secondly,In order to let the model automatically extract features and classification,this paper selects SSD algorithm and YOLO V3 algorithm as the network model for tear break up detection.Analyzed the two model’s network structure,network characteristics and model design principles.In addition,two improvements have been made to the YOLO V3 algorithm.One is to improve the Darknet-53,the backbone network of YOLO V3,and introduce deformable convolution in residual module,which increases the network’s ability to extract irregular deformation features;the second is to use the k-means++ clustering algorithm,so that the clustering results are not affected by the random selection of the initialization center point,and the accuracy of the algorithm is improved.Finally,set up a deep learning experimental environment,import the artificially labeled real frame information into the network model for training,adjust the parameters to obtain the final detection results,and analyze the detection results of tear break up.At the same time,1350 images were selected from 150 samples in the image library for tear break up time detection experiments,and the final results were evaluated.The experimental results show that the three models of SSD algorithm,YOLO V3 algorithm and improved YOLO V3 can complete the detection of tear break up.Among them,the average correct rate of tear break up detection based on the improved YOLO V3 algorithm reaches 92.07%.The accuracy rate is higher,and the missed detection rate is lower.Compared with the SSD algorithm and the YOLO V3 algorithm,it is more suitable for the detection of tear tear break up images.