基于混合深度学习的蔗糖原料蔗顶芽杂质探测技术  

Detection of Impurities in Sugarcane Top Buds Based on Hybrid Deep Learning

作　　者：陈峰练 白琳[1,2] 张扬虎 张茜 李陶深[1,2] CHEN Fenglian;BAI Lin;ZHANG Yanghu;ZHANG Qian;LI Taoshen(School of Computer,Electronics and Information,Guangxi University,Nanning,Guangxi,530004,China;Guangxi Colleges and Universities Key Laboratory of Parallel and Distributed Computing Technology,Nanning,Guangxi,530004,China;The First Affiliated Hospital of Guangxi Medical University,Nanning,Guangxi,530021,China)

机构地区：[1]广西大学计算机与电子信息学院,广西南宁530004 [2]广西高校并行与分布式计算技术重点实验室,广西南宁530004 [3]广西医科大学第一附属医院,广西南宁530021

出　　处：《广西科学》2021年第3期242-248,共7页Guangxi Sciences

基　　金：国家自然科学基金项目(61966003);广西自然科学基金项目(2020GXNSFAA159171)资助。

摘　　要：为有效地解决复杂情况下蔗糖原料蔗顶芽杂质的识别问题,本研究对复杂环境下高密集度的原料蔗顶芽杂质的特征进行分析,将特征金字塔网络(Feature Pyramid Network,FPN)和RoIAlign整合到Faster R-CNN和Cascade R-CNN的架构中,再将多种深度网络算法融入到统一的多层架构的深度网络框架中,创新地设计出一种混合深度学习模型。通过实验验证,在原料蔗顶芽杂质的探测任务中,相较于传统的深度网络模型,本研究提出的混合深度学习模型在性能上有较大的提升。表明本研究提出的混合深度学习模型取得较好的平均精度,其识别性能已经可以达到实际应用水准。In order to effectively solve the problem of identifying sugarcane top bud impurities in sucrose raw materials under complex conditions,this study analyzes the characteristics of high-density raw sugarcane impurities in complex environments,and integrates Feature Pyramid Network(FPN)and RoIAlign into the architecture of Faster R-CNN and Cascade R-CNN.Then a variety of deep network algorithms are integrated into the unified multi-layer deep network framework,and a hybrid deep learning model is innovatively designed.The experimental verification shows that the hybrid deep learning model proposed in this study has a great improvement in performance compared with the traditional deep network model in the detection task of sugarcane top bud impurities.It demonstrates that the hybrid deep learning model proposed in this study has achieved good average accuracy,and its recognition performance has reached the level of practical application.

关 键 词：深度学习 目标检测 农业智能化 甘蔗杂质探测 蔗糖 智能化生产 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]