深度学习结合快速导向滤波识别自然环境下脐橙果实  

Identification of Navel Orange Fruits in Natural Environments Based on Deep Learning Combined with Fast Guided Filtering

作　　者：熊正午 吴瑞梅[2] 黄俊仕 李霸聪 戴仕明[1] 艾施荣[1] XIONG Zhengwu;WU Ruimei;HUANG Junshi;LI Bacong;DAI Shiming;AI Shirong(College of Software,Jiangxi Agricultural University,Nanchang 330045,China;College of Engineering,Jiangxi Agricultural University,Nanchang 330045,China)

机构地区：[1]江西农业大学软件学院,江西南昌330045 [2]江西农业大学工学院,江西南昌330045

出　　处：《江西农业大学学报》2022年第3期736-746,共11页Acta Agriculturae Universitatis Jiangxiensis

基　　金：国家自然科学基金项目(31460315);江西省重点研发计划项目(2017ACF60004)。

摘　　要：【目的】为提高脐橙采摘机器人在自然环境下对脐橙果实进行识别定位的精度,提出采用深度学习结合快速导向滤波方法识别自然环境下脐橙果实。【方法】以赣南脐橙为对象,改进导向滤波方法,去除自然环境下脐橙图像的光照等噪声信息,突出图像颜色和纹理特征。采用带有残差模块的Darknet-53作为特征提取网络,将多尺度融合的3尺度检测网络减少为2尺度检测网络,引入GIoU边界损失函数代替原损失函数,并使用DBSCAN+Kmeans聚类算法,对训练数据集聚类分析,优化预测分支的先验框尺寸,通过迁移学习训练方法建立脐橙果实识别模型,设计单果、向光、背光、果实重叠、枝叶遮挡5组测试集的对比实验,并与其他几种识别模型性能进行比较。【结果】快速导向滤波方法能很好地去除脐橙图像在自然环境下光照及边缘模糊等噪音信息。当优化2个预测分支先验框尺寸时,改进后模型在5种环境下综合性能都优于其他网络,尤其在真实种植环境下识别准确率达到了91.22%,召回率为97.30%,F1平均值为94.16%,识别速率约为26.48 fps。【结论】使用快速导向滤波结合深度学习方法建立的脐橙果实识别模型对自然环境下脐橙果实的识别具有较高的鲁棒性和实时性,为脐橙采摘机器人的视觉识别提供了技术支持。[Objective]In order to improve the recognition and orientation accuracy of the navel orange fruit in the natural environment by the picking robot,this study proposes a deep learning combined with fast guided filtering method to identify navel orange fruit in natural environment.[Methods]This study selected the Gannan navel oranges as the research material.In order to highlight the color and texture characteristics of the images,the improved guided filtering method was used to remove the noise information such as the illumination of the navel orange image in the natural environment.The Darknet-53 with residual module was used as the feature extraction network,the multi-scale fusion 3-scale detection network was reduced to a 2-scale detection network,the GIoU boundary loss function was introduced to replace the original loss function,the DBSCAN+Kmeans clustering algorithm was used to cluster the training data set,and the anchor box size of the prediction branches were optimized.The recognition model for navel orange fruit was established by the transfer learning training method,five test sets of single fruit,light,backlight,overlap and leaf occlusion were designed for the comparison experiment,and the performance was compared with that of other recognition models.[Results]The fast guided filtering method could remove the noise information of the navel orange image in the natural environment,such as illumination and edge blur.When the sizes of three prediction branches were optimized,the comprehensive performance of the improved model was better than other networks in the five environments,especially in the real planting environment,the recognition accuracy rate reached 91.22%,and the recall rate was97.30%,the average value of F1-score was 94.16%,and the recognition rate was about 26.48 fps.[Conclusion]The results show that the navel orange fruit recognition model established by fast guided filtering combined with deep learning methods has higher robustness and real-time performance in the recognition of navel orang

关 键 词：脐橙果实 深度学习 导向滤波 YOLO v3 

分 类 号：S126[农业科学—农业基础科学] S666.4