EfficientNet V2算法融合GCN和CA-Transformer的腐烂草莓分类方法  

Rotten strawberry classification based on EfficientNet V2 algorithm fused with GCN and CA-Transformer

作　　者：王伟 杨世忠 宫钰程 高升 邓兆鹏 WANG Wei;YANG Shizhong;GONG Yucheng;GAO Sheng;DENG Zhaopeng(School of Information and Control Engineering,Qingdao University of Technology,Qingdao,Shandong 266520,China;Shandong Rongxin Aquatic Food Group Co.,Ltd.,Rizhao,Shandong 276800,China;School of Marine Science and Bioengineering,Qingdao University of Science and Technology,Qingdao,Shandong 266042,China)

机构地区：[1]青岛理工大学信息与控制工程学院,山东青岛266520 [2]山东荣信水产食品集团股份有限公司,山东日照276800 [3]青岛科技大学海洋科学与生物工程学院,山东青岛266042

出　　处：《食品与机械》2024年第12期81-88,共8页Food and Machinery

基　　金：国家自然科学基金(编号:62001263);山东省青年基金(编号:ZR2023QC114)。

摘　　要：[目的]利用现代计算机视觉技术和深度学习方法,提升腐烂草莓分类的准确性和效率。[方法]提出了基于EfficientNet V2融合GCN和CA-Transformer的腐烂草莓分类方法。为基准模型添加了图卷积分支,通过聚合节点的周围信息来更新特征表示,更好地捕捉节点在图结构中的上下文信息;将带有注意力的Transformer结构融合到基准模型的主干中,用该结构替换部分卷积操作,实现全局和局部特征的融合,从而更好地识别草莓的腐烂情况;在传统残差结构的基础上引入学习参数,以实现特征的动态融合。[结果]GC-EfficientNet V2模型相比基准模型在准确率上提高了1.86%,召回率提升了1.49%。与Inception V3、ResNet50、VGGNet、Vision Transformer和EfficientNet V2-m模型相比,该模型的识别准确率分别提高了0.93%,2.08%,2.79%,3.26%,0.47%。[结论]该模型能够准确地对腐烂草莓进行分类。[Objective]Improving the accuracy and efficiency of rotting strawberry classification using modern computer vision techniques and deep learning methods.[Methods]A classification method for rotten strawberries based on EfficientNet V2 fusion with Graph Convolutional Network(GCN)and Channel-Attention Transformer(CA-Transformer)has been proposed.Firstly,a graph convolution branch was added to the baseline model,which updated feature representations by aggregating the surrounding information of nodes,better capturing the contextual information of nodes in the graph structure.Secondly,this study integrated the Transformer structure with attention into the backbone of the baseline model,replacing some convolution operations with this structure to achieve the fusion of global and local features,thereby better identifying the rottenness of strawberries.Finally,learning parameters were introduced on the basis of the traditional residual structure to achieve dynamic feature fusion.[Results]The GC-EfficientNet V2 model improved the accuracy by 1.86%and the recall by 1.49%compared to the baseline model.Compared with Inception V3,ResNet50,VGGNet,Vision Transformer,and EfficientNet V2-m,the recognition accuracy of the model was improved by 0.93%,2.08%,2.79%,3.26%,and 0.47%,respectively.[Conclusion]This model can accurately classify rotten strawberries,providing some theoretical support for automatic strawberry sorting.

关 键 词：草莓 腐烂 图卷积 CA-Transformer 可学习残差 EfficientNet V2 

分 类 号：S66[农业科学—果树学]