基于混合注意力机制的香菇菌棒成熟度分级研究  

Maturity Grading of Shiitake Mushroom Logs Based on Bottleneck Attention Module

作　　者：王鲁[1] 王明振 吴秋兰[2] WANG Lu;WANG Ming-zhen;WU Qiu-lan(School of Artificial Intelligence/Shandong Women's University,Jinan 250300,China;College of Information Science and Engineering/Shandong Agricultural University,Tai'an 271018,China;Institute of Big Data/Taishan College of Science and Technology,Tai'an 271038,China)

机构地区：[1]山东女子学院人工智能学院,山东济南250300 [2]山东农业大学信息科学与工程学院,山东泰安271018 [3]泰山科技学院大数据学院,山东泰安271038

出　　处：《山东农业大学学报(自然科学版)》2025年第1期49-57,共9页Journal of Shandong Agricultural University:Natural Science Edition

基　　金：山东省重点研发计划(重大科技创新工程)项目(2022CXGC010609)。

摘　　要：香菇是我国最主要的食用菌品种,实现对香菇菌棒成熟度准确分级是提高香菇产量的前提。本文以香菇菌棒为研究对象,提出了一种基于混合注意力机制的香菇菌棒成熟度分级模型。首先基于采集的香菇菌棒成熟度图像,通过DCGAN模型对香菇菌棒图像进行数据增强,学习各个阶段香菇菌棒成熟度图像的特征分布,构建香菇菌棒成熟度数据集。将混合注意力模块BAM添加到采用分组卷积的ResNeXt网络中,通过自适应调整特征关注重点,产生有效感受,提高分级精度。对比实验分析部分,首先评估了数据增强方法对分级模型的影响,结果表明,本文通过DCGAN模型所构建的数据集在分级中具有更强的鲁棒性,最后将该模型与ResNeXt、VGG-16、ResNet-50在香菇菌棒成熟度数据集中进行对比实验,BAM-ResNeXt模型的准确率、精确率、召回率分别为97.88%、94.26%、97.45%,均优于上述模型,实验结果表明本文提出的BAM-ResNeXt模型在香菇菌棒成熟度分级方面取得了良好的效果。Shiitake mushroom is the most important edible mushroom species in China,and realizing accurate grading of maturity of shiitake mushroom sticks is a prerequisite for improving shiitake yield.In this paper,a maturity grading model of shiitake mushroom sticks based on hybrid attention mechanism is proposed with shiitake mushroom sticks as the research object.Firstly,based on the captured maturity images of shiitake mushroom sticks,the data enhancement of shiitake mushroom stick images is carried out by DCGAN model to learn the feature distribution of the maturity images of shiitake mushroom sticks at each stage,and the maturity dataset of shiitake mushroom sticks is constructed.The hybrid attention module BAM is added to the ResNeXt network using grouped convolution to generate effective feelings and improve the grading accuracy by adaptively adjusting the focus of feature attention.In the section of comparative experimental analysis,the effects of data enhancement methods on the grading model are first evaluated,and the results show that the dataset constructed by the DCGAN model in this paper has stronger robustness in grading,and then the effects of two optimizers,SGD and Adam,as well as the effects of different initial learning rates on the model are comparatively analyzed,and the Adam optimizer and the learning rate of 0.001 are selected as the proposed The parameter settings of the BAM-ResNeXt model,and finally the model was compared with ResNeXt,VGG-16,and ResNet-50 in the mushroom stick maturity dataset in a comparison experiment,and the accuracy,precision,and recall of the BAM-ResNeXt model were 97.88%,94.26%,and 97.45%,which were better than the above models,and the experimental The results show that the BAM-ResNeXt model proposed in this paper has high effectiveness in grading the maturity of shiitake mushroom sticks.

关 键 词：香菇菌棒 成熟度分级 深度学习 注意力机制 生成式对抗网络 

分 类 号：S126[农业科学—农业基础科学]