卷积神经网络在脑脊液图像分类上的应用  

Cerebrospinal fluid images classification based on convolution neural network

作　　者：龚震霆 陈光喜[1] 曹建收 

机构地区：[1]桂林电子科技大学计算机与信息安全学院广西高校图像图形智能处理重点实验室,广西桂林541004

出　　处：《计算机工程与设计》2017年第4期1056-1061,共6页Computer Engineering and Design

基　　金：广西学位与研究生教育改革和发展专项课题基金项目(JGY2014060);广西数字传播与文化软实力中心开放基金项目(ZFZD1408008);国家自然科学基金项目(61462018);广西高校图像图形智能处理重点实验室开放基金项目(LD15042X)

摘　　要：针对脑脊液细胞图像拓扑结构复杂,采用传统的基于人工特征的分类方法效果并不好,提出一种基于卷积神经网络的脑脊液细胞图像分类方法。设计一个网络,卷积层分别使用ReLU、LReLU和RReLU这3种激活函数,分为3个网络模型;CNN-RReLU模型使用RReLU激活函数时采用新的策略,在训练和测试阶段,参数a值都是随机取自区间为5到8的均匀分布。在正常异常脑脊液细胞图像和3类单目标脑脊液细胞图像上的两组实验结果表明,该方法在平均分类准确率标准上有显著提升,单张平均分类时间大幅减少,CNN-RReLU的性能最优,验证了方法的有效性,具有较好的应用价值。For the problem that cerebrospinal fluid cell image＇s topological structure is complex and the classification effect of the traditional method based on artificial feature is poor, an image classification method for cerebrospinal fluid cell based on convolu tion neural network was proposed. A network was designed using different activation functions. ReLU, LReLU and RReLU in the convolution layer were used, which were divided into three models. Among them, when using RReLU activation functions, the CNN-RReLU model adopted a new strategy. In the training and testing phase, variable a was randomly taken from uniform distribution from 5 to 8. Experimental results on two groups of cerebrospinal fluid cell images show that the proposed method has significant improvement on the average classification accuracy standard and the average classification time of an image is greatly reduced, CNN-RReLU has the best performance, which verifies the validity of the proposed method and that the pro- posed method has good application value.

关 键 词：卷积神经网络 脑脊液细胞 图像分类 RReLU激活函数 CNN-RReLU模型 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]