학술논문
改进的VGG16在水稻稻瘟病图像识别中的应用 / Application of improved VGG16 in rice blast image recognition
Document Type
Academic Journal
Author
胡骏; 陆兴华; 林柽莼; 陈嘉铧; 邓雨铮; 许丽娟; HU Jun; LU Xinghua; LIN Chengchun; CHEN Jiahua; DENG Yuzheng; XU Lijuan
Source
计算机应用 / Journal of Computer Applications. 43(z2):196-200
Subject
Language
Chinese
ISSN
1001-9081
Abstract
针对水稻稻瘟病中人工识别的效率低、识别能力差和识别技术应用不普及的问题,提出基于改进VGG16模型的稻瘟病病症精准识别模型——VGG16-H.首先,建立水稻稻瘟病的病斑图像RiceLeafs数据集,利用计算机视觉和OpenCV将RiceLeafs原始数据进行随机旋转、随机亮度变换、随机对比度等操作,以扩充样本数和增强数据;其次,在传统VGG16模型的基础上,减少卷积核数,增加Dropout层和GN(Group Normalization)层,以减少模型参数,降低运算负荷,提高检测性能,加快模型收敛;最后,通过PyTorch深度学习平台训练,使用卷积神经网络(CNN)构建VGG16-H模型.实验结果表明,VGG16-H模型的训练识别率比支持向量机(SVM)和VGG16模型分别提高了2.4和0.8个百分点,测试识别率分别提高了2.4和1.6个百分点.验证了VGG16-H模型能在计算资源有限、水稻病病斑分散条件下提高模型的识别率且不增加过多的训练时耗,在实际农业运用中具有较好的效果.
To address the problems of low efficiency,poor recognition ability and not universal application of recognition technology in manual recognition for rice blast,a precise recognition model for rice blast based on an improved VGG16 model which called VGG16-H was proposed.Firstly,the RiceLeafs dataset of rice blast spot images was built,and the original RiceLeafs data was expanded and enhanced through random rotation,random brightness transformation,radom contrast and other processing operations by using computer vision and OpenCV.Then,based on the traditional VGG16 model,the number of convolutional kernels of the VGG16 model was reduced,and the Dropout layer and GN(Group Normalization)layer were added to reduce the model parameters,reduce the computational load,improve the detection performance and accelerate the model convergence.Finally,the VGG16-H model was constructed by using Convolutional Neural Network(CNN)trained on the PyTorch deep learning platform.The experimental results show that the training recognition rate of the VGG16-H model is 2.4 and 0.8 percentage points,and the test recognition rate is 2.4 and 1.6 percent points,higher than that of Support Vector Machine(SVM)and VGG16 models.The VGG16-H model is proven to be effective in practical agricultural applications,as it can improve the recognition rate without increasing the training time under the conditions of limited computational resources and scattered rice spots.
To address the problems of low efficiency,poor recognition ability and not universal application of recognition technology in manual recognition for rice blast,a precise recognition model for rice blast based on an improved VGG16 model which called VGG16-H was proposed.Firstly,the RiceLeafs dataset of rice blast spot images was built,and the original RiceLeafs data was expanded and enhanced through random rotation,random brightness transformation,radom contrast and other processing operations by using computer vision and OpenCV.Then,based on the traditional VGG16 model,the number of convolutional kernels of the VGG16 model was reduced,and the Dropout layer and GN(Group Normalization)layer were added to reduce the model parameters,reduce the computational load,improve the detection performance and accelerate the model convergence.Finally,the VGG16-H model was constructed by using Convolutional Neural Network(CNN)trained on the PyTorch deep learning platform.The experimental results show that the training recognition rate of the VGG16-H model is 2.4 and 0.8 percentage points,and the test recognition rate is 2.4 and 1.6 percent points,higher than that of Support Vector Machine(SVM)and VGG16 models.The VGG16-H model is proven to be effective in practical agricultural applications,as it can improve the recognition rate without increasing the training time under the conditions of limited computational resources and scattered rice spots.