摘要
从网上下载3种花的图片,利用基于Caffe的卷积神经网络进行分类,比较了5种深度神经网络的分类效果。证明了在这种情况下通过对原始训练样例采用平移技术来增加训练样例数能显著地提高了分类效果,分类准确率提高11%,损失函数从0.88降到0.57,Kappa值从0.63升到0.75。
The pictures of three kinds of flowers are downloaded from Internet and classified with convolutional neural networks based on Caffe.The classification results of five deep neural networks are compared.It is proved that in this situation the technique to generate more train examples by cropping the existing ones could improve the result notably.Classification accuracy increases by11%,the loss function decreases from0.88to0.57and Kappa value increases from0.63to0.75.
作者
詹劲松
ZHAN Jinsong(Fuqing Branch of Fujian Normal University,Fuqing,Fujian 350300)
出处
《福建师大福清分校学报》
2017年第5期42-45,共4页
Journal of Fuqing Branch of Fujian Normal University
基金
福建省教育厅A类项目(JA14341)
