基于特征子图位平面切割和卷积神经网络的青光眼图像分类  

Glaucoma Image Classification Based on Feature Sub-Graph Bit-Plane Slicing and Convolution Neural Network

作　　者：李振鹏 陈国明[1] 孙艳宁 魏小花 李云蓉 Li Zhenpeng;Chen Guoming;Sun Yanning;Wei Xiaohua;Li Yunrong(School of Computer Science,Guangdong University of Education,Guangzhou 510303)

机构地区：[1]广东第二师范学院计算机学院,广州510303

出　　处：《现代计算机》2022年第11期67-73,共7页Modern Computer

基　　金：广东省自然科学基金项目(2018A0303130169);广东省大数据分析与处理重点实验室开放基金项目(201902)。

摘　　要：青光眼是常见的眼科疾病之一,不及时治疗则会导致眼睛失明,而早期青光眼的诊断主要依赖于眼科医生的经验判断,容易出现误诊或漏诊的情况。为了降低该类情况发生的可能性,本文提出了一种利用位平面切割和特征子图技术来提取图像特征,并且结合卷积神经网络(CNN)分类器来提高青光眼识别率的方法。同时为了验证上述方法对提高图像分类准确率的有效性和对防御对抗样本攻击的有效性,本文首先采用快速梯度符号方法(Fast Gradient Sign Method,FGSM)对原始数据集进行攻击,并利用VGG16网络对受到攻击后的图像提取特征子图,然后将特征子图分解为8个位平面图像,最后将位平面图像输入CNN分类器。实验结果表明,运用该方法提取图像特征后,其对应位平面图像的分类准确率相较于受攻击后的图像有显著提高,准确率最高可从66.5%提升到97.4%,提高了30.9%。本文所提出的特征提取方法能有效提高分类准确率,并且具有抵抗对抗样本攻击的能力。Glaucoma is one of the common ophthalmic diseases.If not treated in time,it will lead to blindness.Early diagno⁃sis of glaucoma mainly depends on the experience and judgment of ophthalmologists,which is prone to misdiagnosis or missed diag⁃nosis.In order to reduce this kind of situation,in this paper we propose an image feature extraction method by bit plane slicing,fea⁃ture sub-graph technology and convolutional neural network(CNN)classifier to improve the recognition rate of glaucoma.At the same time,in order to verify the effectiveness of the above methods on improving the accuracy of image classification and defending against adversarial attacks at first we use,the fast gradient sign method(FGSM)to attack the original data set,then we use the vgg16 network is used to extract the feature sub-graph from the attacked image and the corresponding feature sub-graph is decom⁃posed into 8 bit plane images,Finally we put the bit plane image is input into CNN classifier.The experimental results show that the classification accuracy of the corresponding bit plane image is significantly higher than that of the attacked image,which is in⁃creased from 66.5%to 97.4%,and we got 30.9%improvement.The feature extraction methods proposed in this paper can effec⁃tively improve the classification accuracy,and has the ability to resist adversarial attacks.

关 键 词：青光眼 位平面分解 特征子图 快速梯度符号方法 卷积神经网络 

分 类 号：R775[医药卫生—眼科] TP183[医药卫生—临床医学] TP391.41[自动化与计算机技术—控制理论与控制工程]