基于PCANet和SVM的病变眼底图像检测算法

针对眼底图像训练数据集少的问题,该文采用了无监督的主成分分析网络(principal components analysis networks,PCANet)和有监督的支持向量机(support vector mochine,SVM)相结合的算法,通过对彩色眼底图像视网膜渗出物特征的提取,检测出含渗出的糖尿病性视网膜病变眼底图像和正常眼底图像.在对眼底图像进行渗出物特征提取之前,为了减少对渗出物特征提取的干扰,首先对眼底图像进行图像预处理,包括去除冗余背景、通道分离、直方图均衡化、血管去除和视盘去除.无监督的PCANet不需要进行标签训练,与SVM结合,既节约了训练时间,又在训练数据集较小的情况下实现眼底图像的准确分类.实验结果表明:PCANet和SVM相结合的模型在准确性、灵敏度和特异值3个方面与相关方法比较都具有一定的提升.

The Detection Algorithm of Pathological Fundus Image Based on PCANet and SVM

To address the problem of small training data set of fundus images,the combination of unsupervised principal components analysis networks(PCANet)and supervised support vector mochine(SVM)algorithm is used to detect diabetic retinopathy fundus images containing exudates and normal fundus images by extracting retinal exudate features from color fundus images.Before performing exudate feature extraction on fundus images,image preprocessing is first performed on fundus images to reduce interference with exudate feature extraction,including redundant background removal,channel separation,histogram equalization,vessel removal,and optic disc removal.The unsupervised PCANet does not require labels for training and is combined with SVM to both save training time and achieve accurate classification of fundus images with a small training data set.The experimental results show that the PCANet+SVM model has a certain improvement in accuracy,sensitivity and specificity value compared with related methods.