The Application Research On The Recognition Of The Hard Exudate Retinopathy Image

In recent years,the incidence of diabetes show an increasing trend,and diabetic retinopathy as a serious complication of diabetes has received much attention.Besides,human have many retinal blood vessels and the duration of diabetes,serious degree and postoperative recovery were closely associated with its own change.Now,whether the patient is suffering from diabetic retinopathy,they must be manually examined by a professional ophthalmologist.A large number of fundus images,heavy task,heavy intensity and big subjectivity make a certain problem of the diagnosis,coupled with high check the pressure,making the patient can not get timely treatment.So use of computer-based image recognition assistant detection technology has been obtained rapidly development.Automatic detection technology based on image recognition is a hot spot in the field of medical diagnosis.However,there are a lot of interference from image quality and image content in fundus image feature automatic recognition.The automatic identification is often difficult.Aiming at this problem,this paper presents a digital image recognition method based on machine learning and carries out research work.The main contents and contributions of this paper are as follows::Firstly,preprocessing images based on digital image processing technology.we make the color space conversion,image denoising and image enhancement work to eliminate the images of the existence of uneven lighting,low contrast and other issues.After this work,we will improve the overall image quality.Because some visual characteristics of the optic disc and hard exudation is similar,the one-dimensional maximum entropy segmentation algorithm is used to effectively locate and segment the optic disc in order to eliminate the interference of the optic disc in the fundus image.Secondly,for the pre-processed images,in order to distinguish the pixels of the exudation region and the image pixels in the normal region,we extract a significant feature with reference to the recommendations of the clinicians,such as brightness,color,boundary length,etc.,using the mean,the standard deviation and other mathematical statistics indicators to their statistics.And we through the algorithm to construct some new features,such as clustering cluster average brightness.Last,we mainly use two algorithms to identification of hard exudates.One is the Na?ve Bayesian theorem which is based on the independent hypothesis of the characteristic and Bayesian algorithm,the other is the linear classifier support vector machine algorithm with the largest interval in the feature space.Under the jupyter notebook development environment and based on the diaretdb1 database,we first use the naive Bayesian algorithm and the backward wrapping feature selection method to select feature.And then we make these features as support vector machine algorithm initialization feature space,and gradually add features,get ten features as the best feature at this time subset.Finally,the two algorithms are compared with the performance.In addition,this paper compares the recognition results of the above two algorithms,and calculates the accuracy,sensitivity and specificity respectively.The experimental results show that the support vector machine is more effective and has good robustness.