基于YOLON网络的多形态油茶果实夜间检测方法研究

Nighttime detection method of polymorphic Camellia oleifera fruits based on YOLON network

【目的】提出了一种YOLON目标检测网络,为油茶果采收装置夜间非结构化环境下果实目标的精准识别提供技术支持。【方法】在改进YOLOv3的基础上建立YOLON目标检测网络,先在图像输入端添加照度调整模块(LA)对输入图像的照度进行自适应调整,以加强前景图像特征的显著程度,利用特征提取网络对输入图像进行多次卷积以得到对应的特征图;然后在特征融合层添加夜间隐性知识模块(NPK),以先验信息的形式辅助网络预测,提高夜间果实目标的识别准确性;最后对网络特征图进行解码处理得到对应的目标检测框,从而完成对夜间油茶果实目标的检测。为验证所提出网络的有效性,采用准确率(P)、召回率(R)、平均精度均值(mAP)和综合评价指标(F1)对YOLON及对比网络YOLOv3、YOLOv4、YOLOv5s的检测效果进行定量评价。【结果】用YOLON和各对比网络在夜间油茶果数据集上进行训练和测试,YOLON网络的P、R、mAP、F1分别为94.00%,83.63%,94.37%和89.00%,mAP分别较YOLOv3、YOLOv4、YOLOv5s提高2.32%,4.93%和2.33%;对不同果实数量油茶果图像进行测试,YOLON在单果、双果和多果测试数据集上均有较好表现,其对这3类果实目标检测的mAP为98.34%,分别较YOLOv3、YOLOv4、YOLOv5s提高2.17%,8.99%和4.35%;对整树小尺寸油茶果实的检测效果,YOLON的mAP可达93.56%,分别较YOLOv3、YOLOv4、YOLOv5s高1.24%,8.66%和5.57%;在对整树油茶果实图像进行检测时,YOLON的平均置信度为0.69,分别较YOLOv3、YOLOv4、YOLOv5s高0.09,0.22和0.14;此外,用YOLON对夜间采集的处于重叠、遮挡、复杂背景等多态耦合下的油茶果实图像进行检测,也均具有较高的检测置信度。【结论】YOLON可以满足油茶果采收机器人果实定位精度的要求,将其应用于油茶果夜间图像的检测是可行的。

【Objective】A YOLON target detection network was proposed to provide technical support for the accurate recognition of fruit targets in unstructured environment of Camellia oleifera fruit harvesting device at night.【Method】Based on the improvement of YOLOv3,a YOLON target detection network was established.First,a light adjuster(LA)module was added at the image input terminal to adaptively adjust the illumination of the input images and enhance the significance of foreground image features,so as to improve the ability of extracting the features of C.oleifera fruit target at nighttime.The feature extraction network was used to conduct multiple convolutions of the input images to obtain the corresponding feature map.Then,a night prior knowledge(NPK)module in feature fusion layer was added to assist network prediction in the form of prior information for improving the recognition accuracy of night fruit targets.Finally,the network feature map was decoded to obtain the corresponding target detection frame and complete the target detection of C.oleifera fruits at nighttime.The accuracy(P),recall(R),mean average precision(mAP)and comprehensive evaluation index(F1)were used to quantitatively evaluate the detection results of YOLON and comparison networks of YOLOv3,YOLOv4 and YOLOv5s.【Result】YOLON and comparison networks were used for training and testing on the nighttime C.oleifera fruit data set,and the obtained P,R,mAP and F1 of YOLON were 94.00%,83.63%,94.37% and 89.00%,respectively.The mAP of YOLON was 2.32%,4.93% and 2.33% higher than that of YOLOv3,YOLOv4 and YOLOv5s,respectively.YOLON showed good performance in single fruit,double fruits and multiple fruits test data sets with mAP values of 98.34%,which was 2.17%,8.99% and 4.35%higher than that of YOLOv3,YOLOv4 and YOLOv5s,respectively.For whole trees of small C.oleifera fruits,mAP of YOLON reached 93.56%,which was 1.24%,8.66% and 5.57% higher than that of YOLOv3,YOLOv4 and YOLOv5s,respectively.For detecting whole tree of C.oleifera fruit image,average confidence of YOLON was 0.69,which was 0.09,0.22 and 0.14 higher than that of YOLOv3,YOLOv4 and YOLOv5s,respectively.YOLON also had high detection confidence when detecting C.oleifera fruit images collected at night under the polymorphic coupling of overlap,occlusion and complex background.【Conclusion】YOLON could meet the requirements of fruit positioning accuracy of C.oleifera fruit harvesting robots.It was feasible to apply YOLON to detect night images of C.oleifera fruit.