文献类型：期刊文献

中文题名：基于改进YOLOv7的番茄识别和检测算法

英文题名：Tomato recognition and detection algorithm based on improved YOLOv7

作者：罗小娟[1];罗丁楠[1];王兵冰[1];刘柯达[1];吕舒轩[1]

机构：[1]华东理工大学信息科学与工程学院,上海200237

年份：2025

卷号：50

期号：6

起止页码：1209

中文期刊名：广西大学学报(自然科学版)

外文期刊名：Journal of Guangxi University(Natural Science Edition)

收录：;北大核心:【北大核心2023】；

基金：国家自然科学基金项目(61872143);国家级大学生创新实践项目(202310251059)。

语种：中文

中文关键词：深度学习;成熟度判别;注意力机制;激活函数

外文关键词：deep learning;maturity discrimination;attention mechanism;activation function

摘要：针对传统番茄人工采摘易造成品质下降、种植密度大与植株遮挡导致特征提取困难,以及传统卷积神经网络结构复杂、小目标特征提取能力弱、未成熟番茄漏检率高等问题,提出一种基于改进YOLOv7的番茄目标识别和检测模型(YOLOv7-PSF),引入简单无参数注意力模块(SimAM),更有效地捕捉番茄病虫害特征,采用部分卷积(PConv)替代主干网络ELAN结构中CBS模块的Conv以提升计算效率,有效区分被遮挡的番茄,引入激活函数fReLU,减少权值的衰减,解决环境干扰信息与番茄特征信息混淆问题,避免漏检和错检等问题。结果表明:所提出的检测模型的平均精度均值(mAP)为97.4%,模型尺寸为62.7 MB,与原YOLOv7相比,改进后的模型mAP提高了约3.1个百分点,成熟番茄、未成熟番茄和病虫害番茄的F1值分别提高了3个百分点、4个百分点、3个百分点,与其他主流算法相比,改进模型针对番茄的检测性能有明显优势。可视化结果表明,对YOLOv7算法的改进实现了对复杂环境下的番茄高精度检测,可为高密度种植环境下的番茄采摘作业提供有效技术支持。
To address the problems of quality decline from manual tomato picking,obstacles in feature extraction due to dense planting and occlusion,along with the limitations of traditional CNNs,including their complex structure,poor small-target feature detection,and high rates of missing immature tomatoes,a tomato detection and recognition model(YOLOv7-PSF)based on an improved YOLOv7 was proposed.To more effectively capture characteristics of tomato pests and diseases,the SimAM parameter-free attention mechanism was introduced.Partial Convolution(PConv)was adopted to replace standard convolution in CBS module within ELAN structure of the backbone network,which enhanced computational efficiency and improved the recognition of occluded tomatoes.The fReLU activation function was incorporated to reduce weight attenuation and resolve confusion between environmental interference and tomato features,thereby avoiding missed and false detections.Experimental results show that the proposed model achieves a mAP of 97.4%,with a model size of 62.7 MB.Compared with the original YOLOv7,the improved model exhibits an increase in mAP by approximately 3.1 percentage points.The F1 scores for mature,immature,and diseased tomatoes are raised by 3 percentage points,4 percentage points,and 3 percentage points,respectively.The improved model demonstrates clear advantages over other mainstream algorithms in tomato detection performance.Visualization results confirm that the enhanced YOLOv7 algorithm enables high-precision tomato detection in complex environments,providing effective technical support for automated harvesting in high-density cultivation conditions.