应用轻量化FEB-YOLO模型的荔枝果实动态识别计数方法  

Dynamic Recognition and Counting Method for Litchi Fruit Using Lightweight FEB-YOLO Model

作　　者：李景顺 刘美 孟亚男[1] 韩慧子 LI Jingshun;LIU Mei;MENG Yanan;HAN Huizi(School of Information and Control Engineering,Jilin Institute of Chemical Technology,Jilin 132022,China;School of Automation,Guangdong Institute of Petrochemical Technology,Maoming 525000,China;School of Engineering,The Hong Kong Polytechnic University,Hongkong 999077,China)

机构地区：[1]吉林化工学院信息与控制工程学院,吉林吉林132022 [2]广东石油化工学院自动化学院,广东茂名525000 [3]香港理工大学工程学院,中国香港999077

出　　处：《计算机测量与控制》2025年第2期229-237,261,共10页Computer Measurement &Control

基　　金：国家自然科学基金资助项目(62073091);广东省普通高校重点领域(新一代信息技术)专项(2020ZDZX3042)。

摘　　要：针对大场景自然环境下荔枝存在小目标、重叠和遮挡等特点,提出一种轻量化荔枝检测模型FEB-YOLO;该模型基于YOLOv8在C2f模块中引入PConv替代部分常规卷积以实现轻量化改进,同时融入EMA注意力机制提高算法的特征提取能力;将颈部网络替换为融合P2特征层的BiFPN,增强模型对不同尺寸的跨尺度特征融合;在回归损失函数中引入NWD度量,提高模型对荔枝小目标的学习能力,降低漏检率;经实验测试得到FEB-YOLO模型的P、R、mAP对比原始模型分别提高1.4%、1.6%、1.7%,其参数量和计算量分别降低47.3%和27.1%,改进后模型占用的计算资源更少,同时能够明显提高在复杂环境下的识别精度;为实现果园场景下实时估计荔枝产量,提出了一种高效的荔枝果实动态识别计数方法,通过将FEB-YOLO作为BoT-SORT跟踪器的目标检测器,将FEB-YOLO的识别输出作为BoT-SORT的输入,实现动态视频序列的跟踪计数,最后以实例验证了该方法的有效性和可行性;所得改进模型具有较好的鲁棒性且体积小,可以嵌入到边缘设备中,不仅可用于实时估计荔枝产量,还可用于规划采摘和贮藏,为果园资源分配提供可靠支撑。In response to the characteristics of small targets,overlap,and occlusion of litchi in large natural environments,a lightweight litchi detection model FEB-YOLO is proposed.Based on YOLOv8,partial convolution(PConv)is introduced in the module C2f to replace some conventional convolutions to achieve lightweight improvement,and efficient multi-scale attention(EMA)attention mechanism is integrated to improve the feature extraction capability of the algorithm.The neck network is replaced by bidirectional feature pyramid network(BiFPN)with P2 feature layer to enhance the cross-scale feature fusion of different sizes.Normalized weighted distance(NWD)measurement is introduced into the regression loss function to improve the learning ability of the model for litchi small targets and reduce the missed detection rate.Experimental results show that compared with the original model,the P,R and mAP of the FEB-YOLO model are increased by 1.4%,1.6%and 1.7%,respectively,and its Params and FLOPs are reduced by 47.3%and 27.1%,respectively.The improved model has less computing resources,and can improve the recognition accuracy under complex conditions.In order to estimate litchi yield in real-time in orchard scenarios,an efficient dynamic recognition and counting method for litchi fruit is proposed.By using the FEB-YOLO model as the target detector of the BoT-SORT tracker and the recognition output of the FEB-YOLO model as the input of the BoT-SORT tracker,the dynamic video sequence tracking and counting is realized.Finally,the effectiveness and feasibility of the proposed method are verified through examples.The improved model has the features of good robustness,small size,and embedding in edge equipment,which can not only be used for real-time estimation of litchi yield,but also for planning picking and storage,providing reliable support for orchard resource allocation.

关 键 词：荔枝果实 多目标跟踪 产量预测 轻量化 目标检测 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]