基于YOLOv5s-CBS的番茄成熟度识别方法  

Tomato maturity recognition method based on YOLOv5s-CBS

作　　者：丁杰源 奚小波[1] 张宝峰 瞿济伟 朱正波 史扬杰 张瑞宏[1] DING Jieyuan;XI Xiaobo;ZHANG Baofeng;QU Jiwei;ZHU Zhengbo;SHI Yangjie;ZHANG Ruihong(School of Mechanical Engineering,Yangzhou University,Yangzhou 225127,China)

机构地区：[1]扬州大学机械工程学院,江苏扬州225127

出　　处：《扬州大学学报(自然科学版)》2025年第2期29-37,共9页Journal of Yangzhou University:Natural Science Edition

基　　金：江苏省科技计划现代农业项目(BE2018302);江苏省研究生科研与实践创新计划资助项目(SJCX24_2211);扬州大学“高端人才支持计划”。

摘　　要：为精准检测自然环境下番茄果实的成熟情况,依据番茄在生长过程中的颜色变化,将果实成熟度划分为成熟期、转熟期、转色期和未熟期4个等级。针对相邻番茄果实的颜色区分度不明显以及果实堆叠导致识别精度低等问题,提出一种基于YOLOv5s-CBS的番茄成熟度识别方法。该方法引入坐标注意力(coordinate attention, CA)机制,有效抑制背景干扰,提高识别准确性;采用双向加权特征金字塔网络(bi-directional feature pyramid network, BiFPN)替换原有颈部网络结构,增强模型对重要特征的表达能力;运用柔性非极大值抑制(soft non-maximum suppression, Soft-NMS)提升边界框回归效果,减少高置信度检测框被误删的错误。此外,通过引入新的Alpha-CIoU损失函数,进一步提高目标定位精度,加快模型收敛速度。结果表明,YOLOv5s-CBS模型在测试集上的平均精度均值达到95.8%,模型大小14.11 MB;与其他7种模型相比,YOLOv5s-CBS模型的综合性能指标表现最优。本研究实现了番茄成熟度的准确识别,能够为番茄的智能化采摘、分级等提供有力的技术支撑。To accurately detect the maturity of tomato fruits in the natural environment,the maturity is divided into four grades,namely mature,turningmature,color-transition and immature,according to the color changes during the growth process aiming at the challenges such as reduced color distinction of adjacent fruits and low recognition accuracy caused by fruit overlapping,a tomato maturity recognition method based on YOLOv5s-CBS is proposed.The method introduces coordinate attention(CA)mechanism to effectively suppress background interference and enhance recognition accuracy.At the same time,a bidirectional weighted feature pyramid network(BiFPN)is adopted to replace the original neck network structure to enhance the model's ability to express important features.In addition,a new loss function Alpha-CloU is introduced to further improve the target positioning accuracy and accelerate the convergence speed.To reduce the erroneous deletion of high-confidence detection boxes,soft non-maximum suppression(Soft-NMS)is used to improve the bounding box regression rate.Experimental results demonstrate that the mean average precision(mAP)of the YOLOv5s-CBS model on the test set reaches 95.8%,and the model size is 14.11 MB.Compared with other seven models,the YOLOv5s-CBS model has the best comprehensive performance.This study enables accurate tomato maturity identification,providing robust technical support for intelligent harvesting and grading of tomatoes.

关 键 词：番茄 成熟度识别 注意力机制 损失函数 YOLOv5 

分 类 号：S224.9[农业科学—农业机械化工程] TP751.1[农业科学—农业工程]