基于MobileNets的果园害虫分类识别模型研究
2018-11-28樊景超
樊景超
摘 要:为在小数据集上构建果园害虫分类识别模型,并在普通安卓手机运行,本研究收集了6种常见果园害虫共计3 058张图像,首先对害虫图像进行反卷积遴选,再通过MobileNets构建深度学习模型进行分类识别。结果表明,所构建模型在安卓手机上运行流畅,可对所采集的6种果园害虫进行识别,其准确率均在90%以上,效果良好。
关键词:果园害虫;MobileNets;分类识别
中图分类号:TP311 文献标识码:A DOI 编码:10.3969/j.issn.1006-6500.2018.09.004
Study on Classification and Recognition of Orchard Pests Based on MobileNets
FAN Jingchao
(Agricultural Information Institute, Chinese Academy of Agricultural Sciences, Beijing 100081, China)
Abstract: In order to construct the model of orchard pest classification on the small data set and run identification of model on ordinary Android mobile phone, a total of 3 058 images of 6 common orchard pests were collected. First, the image of the pest was selected by deconvolution, and then the depth learning model was constructed by MobileNets. The constructed model could run well on Android mobile, and identify the selected 6 common orchard pests. The result showed that the recognition accuracy was over 90%.
Key words: orchard pest; MobileNets; classification and recognition
我國是苹果生产大国[1-3],同时又是有害生物危害较为严重的国家之一,苹果生产中有害生物的种类多,发生频率高,且分布地域广[4-11]。害虫分类识别是虫情预测、精准施药等工作的前提基础,在苹果生产、仓储、检疫等领域有着强烈的现实需要,其传统方法依靠专家远程诊断和农技推广人员的识别,劳动强度高、效率低。随着信息技术的兴起,在果园害虫图像分类识别方法上基于纹理特征提取的计算机视觉识别方法成为热点,如直方图、集合不变量、子空间等[12-15],其本质是以专家经验进行人工特征的设计,这类方法一般要求识别环境可控,如将害虫放置在实验室条件下的单一背景中进行特征数据采集,然而在实际应用环境——野外自然条件中,背景和光照以及昆虫姿态均属于不可控因素,大大增加了特征提取的难度。再加上害虫本身纹理丰富、结构复杂,尤其在自然环境下,害虫姿态变化丰富,种内和种间差异大,且背景环境复杂,使图像识别成为一种细粒度识别问题[16],不可避免主观误差,造成所构建模型的适应性较差,技术难度大大增加。……
