摘要
为了解决传统卷积神经网络在黄瓜叶部病斑图像分割中存在模型训练时间长、分割效果差以及分割过程中易受光照和背景影响等问题,提出了一种基于改进全卷积神经网络的黄瓜叶部病斑分割方法。首先在模型训练的初始阶段使用传统的卷积神经网络得到病斑图像的轮廓特征,在训练过程中将传统的修正性单元(RELU)激活函数替换为指数线性单元(ELU)激活函数;然后对传统的卷积神经网络得到的病斑图像轮廓特征进行二次模型训练,训练过程中使用批归一化(Batch normalization)函数稳定模型训练过程;最后将原始卷积神经网络的多项逻辑回归(Soft max)分类器更换为支持向量机(SVM)分类器,对分类器输出的像素分类结果进行反卷积操作,恢复图像分辨率,得到分割结果。使用本研究方法与改进OTSU、SVM、CRF和传统FCN等4种方法在黄瓜叶部病斑数据集上进行分割试验,结果表明本研究方法的平均像素分割准确率为80.46%,平均交并比为70.43%,具有较高的分割精度。
To solve the problems of the traditional convolutional neural network(CNN)in the process of cucumber leaf diseases image segmentation,such as long model training time,poor segmentation effect and easy to be affected by light and background in the process of segmentation,a method for segmentation of cucumber leaf diseases based on improved fully convolutional neural network was proposed.Firstly,in the initial stage of model training,traditional CNN was used to obtain the contour features of the diseases image.In the process of training,activation function of rectified linear units(RELU)was replaced by the exponential linear unit(ELU).Secondly,the disease contour features obtained by the traditional CNN were trained twice,and the batch normalization function was used to stabilize the model training process.Finally,the SoftMax of the original CNN was replaced with support vector machine(SVM),and the pixel classification result outputs by the classifier were deconvolution operation to restore the image resolution and obtain the segmentation results.The segmentation experiment was carried out on the cucumber leaf image by using this research algorithm and others four algorithms including improved OTSU,SVM,CRF and traditional FCN.The results showed that the average pixel segmentation accuracy of this algorithm was 80.46%,and the average intersection ratio was 70.43%,which could accurately segment the diseased parts in the leaves.
作者
王振
张善文
王献锋
WANG Zhen;ZHANG Shan-wen;WANG Xian-feng(College of Information Engineering,Xijing University,Xi′an 710123,China)
出处
《江苏农业学报》
CSCD
北大核心
2019年第5期1054-1060,共7页
Jiangsu Journal of Agricultural Sciences
基金
国家自然科学基金项目(61473237)
关键词
黄瓜病斑图像
卷积神经网络
激活函数
图像分割
cucumber lesion image
convolution neural network
activation function
image segmentation
作者简介
王振(1994-),男,河南周口人,硕士研究生,研究方向为模式识别技术在农业领域的应用。(E-mail)wangzhen4013@163.com。
