提出一种基于超像素仿射传播聚类的视网膜血管分割方法。首先对预处理后的图像提取Hessian最大本征值、Gabor小波、B-COSFIRE滤波特征,构建3维眼底图像像素特征;同时对眼底图像进行超像素分块,并采用一致性准则对所分的超像素块进行筛选,得到超像素候选块;把超像素候选块当作样本点,把候选块内的像素特征的统计平均值当作特征向量,在特征空间中进行仿射传播聚类得出血管类和背景类两个聚类中心;根据血管类和背景类两个聚类中心,采用最近邻方法对眼底像素进行分类,实现对视网膜血管的分割。实验表明:在DRIVE和STARE眼底图像数据库上,本文算法的平均准确率分别为94.63%和94.30%;相较于K-means、模糊C均值(FCM)和其他聚类方法,本方法对血管的识别度高,所分割的视网膜血管有较好的连续性和完整性。

A retinal vessel segmentation method based on the affinity propagation clustering of superpixels was proposed herein. First, the maximum Hessian eigenvalue, the Gabor wavelet, and the B-COSFIRE filtering features were extracted from the preprocessed image to construct the three-dimensional fundus image. The fundus image was segmented into superpixel blocks, which were screened based on a pixel consistency criterion to select the best candidates; these candidates were considered as sample points and their statistical average pixel values were used as the feature vectors. Two clustering centers of the vessel and background classes were obtained by performing affinity propagation clustering on the feature space. Based on these clustering centers, the fundus pixels were classified via the nearest neighbor method for retinal vessel segmentation. The experimental results show that the accuracies are 94.63% and 94.30% for the DRIVE and STARE fundus image databases, respectively. Compared with K-means clustering, FCM (Fuzzy C-means), and other clustering methods, the proposed technique presents a high recognition degree for blood vessels and better continuity and integrity of the segmented retinal vessels.