基于栈式稀疏自编码器的青光眼眼底图像识别研究

Study on Fundus Image Recognition of Glaucoma Based on Stacked Sparse Auto-encoder

青光眼是一种常见的威胁视神经及视觉功能的眼病,其具有发病率高,难以察觉等特点。但是目前对青光眼的识别诊断方法还不是很完善,且识别方法复杂,识别率也不高。因此提出了一种基于栈式稀疏自编码器的眼底图像特征提取及图像识别的方法。该方法采用逐层贪婪训练法从无标签的数据集中学习到数据的内部特征,将学习到的特征作为softmax分类器的输入。然后利用带标签的数据通过反向传播算法对稀疏自编码器进行调优。仿真实验分析中,使用测试集数据对该方法进行验证,精确度可达89%,并且优于实验中的其他方法,对青光眼的识别具有一定的实用价值。

Glaucoma is a common eye disease that attacks optic nerve and visual function,and it has the characteristics ofhigh incidence,difficult to detect and so on. However,the current methods for the diagnosis of glaucoma are not mature,and therecognition methods are complex and the recognition rate is not high. Therefore a method about feature extraction of fundus imageand image recognition based on stacked sparse auto-encoder is proposed. The method uses greedy layer-wise training method tolearn the internal characteristics of data from unlabeled data sets,and uses the learning features as input to the softmax classifier.Then,the sparse encoder is tuned by back propagation algorithm using labeled data. In the simulation experiment,the method isvalidated by using the test set data. The accuracy is up to 89%,and it is better than other methods in the experiment. It is of practi-cal value for the recognition of glaucoma.