摘要
目前多肉植物产业在我国发展较快,市场前景广阔,由于其具有品种繁多、形态多变、类间相似度高等特点,导致多肉植物种类辨别难度较大。针对上述问题,提出一种基于改进MobileNet V3网络与迁移学习的多肉植物图像分类方法,将Bottleneck模块前六层的ReLU激活函数换成LeakyReLU激活函数,优化了SE模块,添加了Dropout层提高模型的泛化性。通过改进MobileNet V3网络对13种多肉植物图像进行种类识别,准确率为97.35%,并且可以实时稳定地对多肉植物图像进行分类,使用Focal Loss代替交叉熵损失函数,达到平衡样本的目的。研究结果表明,利用改进MobileNet V3网络对多肉植物种类识别具有一定可行性。
At present,the succulent industry is developing rapidly in China and has broad market prospects.It is difficult to distinguish the succulent species due to their wide variety,variable shapes,and high similarity between species.For the above problem,a succulent image classification method is proposed based on the improved MobileNet V3 network and transfer learning.The ReLU activation function of the first six layers of the Bottleneck module is replaced with the LeakyReLU activation function,the SE module is optimized,and the Dropout layer is added to improve model generalization.By using improved MobileNet V3 network,13 kinds of succulent images are identified by species,the accuracy rate is 97.35%,and the succulent images can be classified stably in real time.Focal Loss is adopted to replace the Cross entropy loss function to achieve the purpose of samples balancing.Experimental and simulation results show that it is feasible to apply the improved MobileNet V3 network to identify succulent plant varieties.
作者
孙公凌云
张靖渝
连俊博
宁景苑
刘伟立
刘权
王国振
陆诗怡
时鹏辉
楼雄伟
SUN Gonglingyun;ZHANG Jingyu;LIAN Junbo;NING Jingyuan;LIU Weili;LIU Quan;WANG Guozhen;LU Shiyi;SHI Penghui;LOU Xiongwei(School of Mathematics and Computer Science,Zhejiang Agriculture and Forestry University,Hangzhou Zhejiang 311300,China)
出处
《传感技术学报》
CAS
CSCD
北大核心
2023年第12期1916-1927,共12页
Chinese Journal of Sensors and Actuators
基金
浙江省科技计划项目(2019C02075)。
