一种基于深度学习的端到端生菜无损鲜重估测模型的建立

A model for end-to-end non-destructive fresh weight estimation of lettuce based on deep learning

[目的]为解决生菜分类、分割和鲜重估测独立处理,增加时间成本的问题,本文提出一种端到端的生菜无损鲜重估测模型——LettuceNet。[方法]LettuceNet模型通过分析俯视图像估测生菜鲜重。LettuceNet结合Swin Transformer-Tiny(Swin-T)和UPerNet,有效提取生菜冠层图像的特征。模型设计了基于K-Net的用于语义分割的分割头部网络以及用于鲜重估测的回归头部网络。回归头部网络融合利用Swin-T的特征与分割头部网络的结果,用于生菜的分类和冠层面积统计,使LettuceNet能够同时高效处理语义分割和鲜重估测任务。[结果]2个数据集的试验结果表明,LettuceNet语义分割任务中,其平均像素准确度(MPA)分别为98.01%和98.75%,而平均交并比(MIoU)分别为96.02%和97.63%;在鲜重预测方面,决定系数R 2分别为0.898和0.919,均方根误差(RMSE)分别为0.865和30.814 g,平均绝对百分比误差(MAPE)分别为1.894%和18.194%。[结论]通过输入生菜冠层图像,LettuceNet能够实时且无损完成生菜的分类、分割与鲜重估测,能够快速对生菜的生长情况进行定量分析,为植物工厂的智能管控提供数据支持。

[Objectives]To address the issue of increased time cost due to separate processing of lettuce classification,segmentation,and fresh weight estimation,this paper proposed an end-to-end non-destructive fresh weight estimation model for lettuce,called LettuceNet.[Methods]LettuceNet estimated lettuce fresh weight by analyzing top-view images.The model integrated Swin Transformer-Tiny(Swin-T)and UPerNet to effectively extract features from lettuce canopy images.A K-Net-based segmentation head was designed for semantic segmentation,while a regression head was used for fresh weight estimation.The regression head combined the features extracted by Swin-T with the results of the segmentation head for lettuce classification and canopy area measurement,allowing LettuceNet to efficiently handle both semantic segmentation and fresh weight estimation tasks.[Results]Experiments on two datasets showed that LettuceNet achieved mean pixel accuracy(MPA)of 98.01%and 98.75%,and mean intersection over union(MIoU)of 96.02%and 97.63%,respectively,for the semantic segmentation task.For fresh weight estimation,the coefficient of determination(R 2)was 0.898 and 0.919,the mean squared error(RMSE)was 0.865 and 30.814 g,and the mean absolute percentage error(MAPE)was 1.894%and 18.194%,respectively.[Conclusions]By inputting lettuce canopy images,LettuceNet can perform real-time,non-destructive lettuce classification,segmentation,and fresh weight estimation.The model enables rapid quantitative analysis of lettuce growth,providing data support for intelligent management in plant factories.