Research On Blood Vessel Segmentation Method Of Fundus Images

The microvasculature in the posterior part of the human eye,i.e.the fundus,is the only deep blood vessel in the human body that can be directly observed in a non-invasive way.The changes in its morphological structure can reflect eye diseases such as glaucoma,cataract,age-related macular degeneration,etc.At the same time,it is closely related to other body organ diseases such as diabetes,hypertension,arteriosclerosis,etc.In addition,the uniqueness of fundus blood vessel distribution makes it an important biological feature of identity authentication.Therefore,it has important application value and practical significance to study how to accurately identify blood vessels in fundus images.Before the segmentation of fundus images,we need to preprocess the original fundus images.First,extract the green channel image of the fundus image;Secondly,the image pixels of the green channel should be reversed.Because the color of the blood vessel and background is close after inversion,therefore,a fundus image with black background is obtained through the mask image.At the same time,in order to avoid the influence of retinal boundary,we use morphological operation to expand the retinal boundary.After the preprocessing,two unsupervised methods and two supervised methods were used to realize the blood vessel segmentation of fundus images.The specific research contents are as follows:(1)Based on morphological blood vessel segmentation.The blood vessels in the fundus image can be seen as a large number of smaller segments.Therefore,morphological operation can be used to separate the complete vascular network,and the morphological operation has the advantages of high efficiency,fast speed and good noise suppression.(2)A multiscale linear detector for blood vessel segmentation.By changing the length of the basic linear detector,the method can extract the linear structure of the blood vessels from the fundus image well so as to obtain better segmentation results.Although the multi-scale linear detector is an unsupervised segmentation method,it can effectively improve the situation of vascular fusion and fracture.(3)The method of blood vessel segmentation based on 2D Gabor.The method uses the pixel gray value and the results of the 2D Gabor wavelet transform at each scale as the feature vector and uses the Bayesian Gauss mixing classifier to classify the vessel features.(4)A blood vessel segmentation method based on Fully Connected Conditional Random Field(FC-CRF).This method takes into account the relationship between pixels,and the structured output support vector machine(SOSVM)is used to calculate and automatically adjust the weight vector and feature map involved in the segmentation process.It improves the connectivity of blood vessels and contributes to the segmentation of small vessels.Finally,this thesis takes the fundus images in the two open databases DRIVE and HRF as experimental objects to verify and compare the above four methods.The segmentation results show that unsupervised class segmentation method is simple and easy to implement,but its accuracy is lower than that of supervised class segmentation method.Supervisory methods require complicated training and learning processes,which results in the increased amount of computation,but its accuracy of segmentation is improved.