Abstract: Machine learning requires a manually annotated set of training images for classifying a pixel either as a vessel or a non-vessel in previously unseen image. It’s difficult to obtain the training samples and expensive time. A new unsupervised learning approach based on feature fusion is proposed. Firstly, a set of 18-D discriminative feature vectors, consisting of Hu moment invariants, Hessian, phase congruency, Gabor wavelet transform, Candy edge detector, are extracted for each pixel of the fundus image. Then a matrix based on the feature vectors is divided into multimodel sets, and uses the [k]-means method to cluster respectively. Finally, the clustering?results are combined as output of the retinal vessel segmentation. Experimental results show that the proposed approach has good average accuracy and running time.

Key words: retina vessel segmentation, feature combination, multimodel learning, [k]-means

摘要： 有监督的学习方法用于视网膜血管分割须以专家手动标记好的视网膜血管为标准，存在训练样本获取困难且训练时间长等不足。针对这些缺点，提出一种基于特征组合的多模块无监督学习方法，提取眼底图像素的不变矩、Hessian矩阵、相位一致性、Gabor小波变换、Candy边缘共18维特征向量，采用多模块[k]-means方法进行视网膜血管分割。实验结果表明，该方法简单，具有较好的准确度，且时间开销少。

关键词: 视网膜血管分割, 特征组合, 多模块学习, [k]-means