摘要
初烤烟叶等级的快速准确检测对开发烟叶智能分级设备以促进农产品精细化管理有着重要意义。针对相似度较高但等级不同的初烤烟叶难以区分的问题,本文提出多感受野特征自适应融合及动态损失调整的初烤烟叶等级检测网络(Flue-cured Tobacco Leaf Grade Detection Network,FTGDNet)。首先,FTGDNet采用CSPNet作为特征提取主干网络,采用GhostNet作为辅助特征提取网络以增强模型的特征提取能力;其次,在主干网络末端嵌入显式视觉中心瓶颈模块(Explicit Visual Center Bottleneck module,EVCB)以实现全局特征信息与局部细节特征信息融合;然后,构建多感受野特征自适应融合模块(Multi-Receptive Field Feature Adaptive Fusion module,MRFA_d),利用注意力特征融合机制(Attention Feature Fusion,AFF)将不同感受野特征图进行自适应加权融合,在增强模型局部感受野的同时突出有效通道信息;最后,设计了一种新的定位损失函数(More Complete IoU Loss,MCIoU_Loss),结合预测框与真实框面积损失以解决在回归定位过程中二者宽高比相等且中心点重合时CIoU_Loss性能退化导致定位精度下降问题,此外,引入矩形相似度衰减系数在训练过程中对真实框与预测框的相似度判别项进行动态调整,加快模型拟合。实验结果表明,FTGDNet对十个等级的初烤烟叶的验证精度达到90.0%,测试精度达到87.4%,且推理时间仅为12.6 ms。相较于多种先进目标检测算法,FTGDNet具有更高的检测精度和更快的检测速度,可为高精度初烤烟叶等级检测提供关键技术支撑。
Rapid and accurate detection of flue-cured tobacco leaf grade is integral to the advancement of tobacco intelligent equipment,promoting refined management of agricultural products.Aiming at the issue that it is difficult to distinguish flue-cured tobacco leaves with high similarity between different grades,a flue-cured tobacco leaf grade detection network(FTGDNet)through multi-receptive field feature fusing adaptively and dynamic loss adjustment was proposed.Firstly,FTGDNet adopted CSPNet and GhostNet as feature extraction backbone network and auxiliary feature extraction network to enhance the model feature extraction ability,respectively;Secondly,to merge global feature information and local detail feature information,an explicit visual center bottleneck module(EVCB)was embedded at the end of backbone network;Moreover,a multi-receptive field feature adaptive fusion module(MRFA_d)was constructed,in which the attention feature fusion(AFF)mechanism adaptively fuses the weights of feature maps with different receptive fields to highlight the effective channel information while enhancing the local receptive fields of the model;In addressing the decrease of positioning accuracy due to CIoU_Loss performance degradation when the prediction box and real box shared the same aspect ratio and their centers align during the regression positioning process,a new positioning loss function MCIoU_Loss was designed,In addition,the rectangular similarity attenuation coefficient was introduced to dynamically adjust the similarity discriminant of prediction box and real box to accelerate the model fitting.The experimental results show that the verification accuracy and test accuracy of FTGDNet for 10 grades of flue-cured tobacco leaf reached 90.0% and 87.4%,respectively,with an inference time of 12.6 ms.Compared with various advanced object detection network,FTGDNet achieves higher detection accuracy and faster detection speed,which could provide technical support for high-precision flue-cured tobacco leaf grade detection.
作者
何自芬
罗洋
张印辉
陈光晨
陈东东
徐林
HE Zifen;LUO Yang;ZHANG Yinhui;CHEN Guangchen;CHEN Dongdong;XU Lin(Faculty of Mechanical and Electrical Engineering,Kunming University of Science and Technology,Kunming 650500,China)
出处
《光学精密工程》
EI
CAS
CSCD
北大核心
2024年第2期301-316,共16页
Optics and Precision Engineering
基金
国家自然科学基金资助项目(No.62171206,No.62061022)
中国烟草云南分公司烟叶智能分级项目资助(No.HZ2021K0462A)。
关键词
初烤烟叶
目标检测
多感受野特征融合
动态损失调整
flue-cured tobacco leaf
object detection
multi receptive field feature fusion
dynamic loss adjustment
