基于Swin Transformer与GRU的低温贮藏番茄成熟度识别与时序预测研究  

Low Temperature Storage Tomato Maturity Recognition and Time Series Prediction Based on Swin Transformer-GRU

作　　者：杨信廷[1,2] 刘彤 韩佳伟 郭向阳 杨霖 YANG Xinting;LIU Tong;HAN Jiawei;GUO Xiangyang;YANG Lin(Department of Information,Shanghai Ocean University,Shanghai 201306,China;Research Center of Information Technology,Beijing Academy of Agriculture and Forestry Sciences,Bejing 100097,China;National Engineering Laboratory for Agri-product Quality Traceability,Beijing 100097,China;Department of Information Science and Technology,Zhongkai University of Agriculture and Engineering,Guangzhou 510225,China)

机构地区：[1]上海海洋大学信息学院,上海201306 [2]北京市农林科学院信息技术研究中心,北京100097 [3]农产品质量安全追溯技术及应用国家工程研究中心,北京100097 [4]仲恺农业工程学院信息科学与技术学院,广州510225

出　　处：《农业机械学报》2024年第3期213-220,共8页Transactions of the Chinese Society for Agricultural Machinery

基　　金：国家重点研发计划项目(2023YFD2001302、2022YFD2001804);北京市农林科学院科研创新平台建设项目(PT2023-24)。

摘　　要：面向绿熟番茄采后持续转熟特征,适时调温是满足不同成熟度番茄适宜贮运温度需求的关键,而果实成熟度自动识别与动态预测则是实现温度适时调控的基础条件。本文基于Swin Transformer与改进GRU提出了一种番茄成熟度识别与时序动态预测模型,首先通过融合番茄两侧图像获取番茄表观全局红色总占比,构建不同成熟番茄图像数据集,并基于迁移学习优化Swin Transformer模型初始权重配置,实现番茄成熟度分类识别;其次,周期性采集不同储藏温度(4、9、14℃)下番茄图像数据,结合番茄初始颜色特征与贮藏环境信息,构建基于Swin Transformer与GRU的番茄成熟度时序预测模型,并融合时间注意力模块优化模型预测精度;最后,对比分析不同模型预测结果,验证本研究所提模型的准确性与优越性。结果表明,番茄成熟度正确识别率为95.783%,相比VGG16、AlexNet、ResNet50模型,模型正确识别率分别提升2.83%、3.35%、12.34%。番茄成熟度时序预测均方误差(MSE)为0.225,相比原始GRU、LSTM、BiGRU模型MSE最高降低29.46%。本研究为兼顾番茄成熟度实现贮藏温度柔性适时调控提供了关键理论基础。Targeting the continuous ripening process of green mature tomatoes after harvest,timely temperature adjustment plays a pivotal role in meeting the appropriate storage and transportation temperature requirements for tomatoes at different stages of ripeness.Meanwhile,automatic recognition and dynamic prediction of fruit ripeness serve as fundamental prerequisites for achieving temperature control at the right time.A tomato ripeness recognition and temporal dynamic prediction model was proposed based on Swin Transformer and improved GRU.Firstly,by fusing the images of both sides of tomatoes,the overall redness proportion as a visual feature was obtained and a dataset of tomato images at different ripeness stages was constructed.Through transfer learning,the initial weight configuration of the Swin Transformer model was optimized to achieve tomato ripeness classification.Secondly,tomato image data at different storage temperatures(4℃,9℃and 14℃)was periodically collected,and the initial color features of tomatoes were combined with storage environment information to build a tomato ripeness temporal prediction model based on Swin Transformer and GRU.Furthermore,a time attention module was incorporated to enhance the prediction accuracy of the model.Lastly,the prediction results of different models were compared and analyzed to validate the accuracy and superiority of the proposed model.The results demonstrated a correct recognition rate of 95.783%for tomato ripeness classification,with respective improvements of 2.83%,3.35%,and 12.34%compared with that of the VGG16,AlexNet,and ResNet50 models.The mean square error(MSE)for tomato ripeness temporal prediction was 0.225,representing a maximum reduction of 29.46%compared with that of the original GRU,LSTM,and BiGRU models.The research result can provide a key theoretical basis for the flexible and timely regulation of storage temperature considering tomato maturity.

关 键 词：番茄 低温贮藏 成熟度识别 时序预测模型 Swin Transformer GRU 

分 类 号：S641.2[农业科学—蔬菜学]