摘要
[目的]开发基于图像的香花油茶品种识别技术,快速、准确识别香花油茶无性系品种。[方法]选择自然光照条件下生长的20个香花油茶无性系品种的油茶果作为研究对象,利用红米K30 Pro、华为P40、华为nova7、i Phone 12、魅族16s多种品牌型号的智能手机对自然状态下油茶果的脐面、侧面进行图像采集,去除低质量图像,通过数据增强方法增加图像数量,采用深度学习网络GoogLeNet-V3对20个品种的香花油茶果图像进行识别。同时对使用多个设备采集图像的品种,进行不同设备型号及数量、相同设备数量但不同比例等情况可能造成的影响进行探讨。[结果]构建了图像数量为16 832张的香花油茶果品种数据集。使用GoogLeNet-V3模型能满足基于油茶果图像的香花油茶品种识别要求,其中识别准确率、召回率、宏观F1值和微观F1值分别为89.13%、89.31%、89.22%和94.29%。对多设备的研究结果表明数据量的增加能有效提高模型精度,且在保证单设备采集数据量的条件下,使用多种设备采集数据构建的模型具有更高的鲁棒性。以GoogLeNet-V3模型为基础构建的移动端识别APP,具有PC端同等精度,可用于香花油茶果品种的识别。[结论]使用深度学习网络GoogLeNet-V3模型能够实现香花油茶果的品种识别。
【Objective】Develop image-based Camellia osmantha variety recognition technology to quickly and accurately identify asexual varieties of C.osmantha.【Method】Twenty C.osmantha varieties grown under natural lighting conditions were selected as research objects,and images of umbilical and lateral parts of C.osmantha fruits in their natural state were collected with using various brands and models of smartphones such as Redmi K30 Pro,Huawei P40,Huawei nova7,iPhone 12,and Meizu 16s.The low quality images were removed and the number of images was increased by data augmentation methods.The images of the 20 C.osmantha varieties were recognized using GoogLeNet-V3.At the same time,for varieties with images collected by multiple devices,the possible impact of different device models and quantities,the same device quantity but different proportions was discussed.【Result】A dataset of 16832 images of C.osmantha fruit varieties was constructed.The GoogLeNet-V3 met the requirements of C.osmantha varieties recognition based on fruits images;among them,the average accuracy,recall,Macro F1 value and Micro F1 were 89.13%,89.31%,89.22%and 94.29%,respectively.The results of the study on multiple devices showed that the increase of data volume can effectively improve the model accuracy,and the model constructed by using data collected from multiple devices had higher robustness under the condition of ensuring the data volume of single device collection.The mobile recognition APP based on the GoogLeNet-V3 model had the same accuracy as the PC side and could be used for the identification of C.osmantha fruit varieties.【Conclusion】The GoogLeNet-V3 deep learning network model can achieve C.osmantha variety recognition based on fruits.
作者
尹显明
彭邵锋
程俊媛
陈梦秋
张日清
莫登奎
韦维
严恩萍
YIN Xianming;PENG Shaofeng;CHENG Junyuan;CHEN Mengqiu;ZHANG Riqing;MO Dengkui;WEI Wei;YAN Enping(Central South University of Forestry&Technology,Changsha 410004,Hunan,China;Hunan Academy of Forestry,Changsha 410004,Hunan,China;Guangxi Forestry Research Institute,Nanning 530002,Guangxi,China)
出处
《经济林研究》
北大核心
2023年第3期70-81,共12页
Non-wood Forest Research
基金
湖南省林业科技创新专项(XLK201939)
广西壮族自治区林业科学研究院横向课题(GXLKY-15126083)。
关键词
香花油茶
品种识别
深度学习
油茶果
Camellia osmantha
variety identification
deep learning
Camellia fruit
