首页 | 本学科首页   官方微博 | 高级检索  
     检索      

基于双维度注意力机制的西红柿成熟度检测方法
引用本文:赵立新,白银光,何春燕,张程,李雅婧,赵树国.基于双维度注意力机制的西红柿成熟度检测方法[J].科学技术与工程,2023,23(11):4571-4578.
作者姓名:赵立新  白银光  何春燕  张程  李雅婧  赵树国
作者单位:河北工程大学;邯郸职业技术学院
基金项目:河北省重点研发计划项目
摘    要:西红柿的成熟度对于采摘、运输和销售至关重要。针对西红柿需要在特定环境下识别问题,从种植现场拍摄图片来制作数据集,提出了一种基于双维度注意力机制的西红柿成熟度分类方法。首先通过tensorflow搭建卷积神经网络,网络中加入了改进的CBAM(convolutional block attention module)模块提取西红柿的成熟度和所在位置信息,即在通道注意力模块中并行一个共享多层感知器后的平均池化层;然后使用Adam优化器更新参数,此方法不仅缓解了网络中直接加入CBAM模块出现的不稳定问题,而且加快了损失函数的下降速度;最后通过调节学习率并使用混淆矩阵计算验证集的准确率获取最佳模型。实验结果表明,本文所提网络在训练了30个Epoch后损失函数稳定下降,搭建软件测试平台进行测试后得到准确率为99%,单张图片检测时间为1.5 s。检测时间和测试准确率均优于AlexNet网络,Grad-CAM可视化结果显示本文所提网络提取目标信息的效果优于AlexNet网络和改进之前的CBAM模块。本文所提方法适用于任意背景下的瓜果品级分类。

关 键 词:神经网络  西红柿  注意力  学习率
收稿时间:2022/8/22 0:00:00
修稿时间:2023/5/5 0:00:00

Tomato maturity detection based on two-dimensional attention mechanism
Zhao Lixin,Bai Yinguang,He Chunyan,Zhang Cheng,Li Yajing,Zhao Shuguo.Tomato maturity detection based on two-dimensional attention mechanism[J].Science Technology and Engineering,2023,23(11):4571-4578.
Authors:Zhao Lixin  Bai Yinguang  He Chunyan  Zhang Cheng  Li Yajing  Zhao Shuguo
Institution:Hebei University of Engineering
Abstract:The maturity of tomatoes is crucial for picking, transportation and marketing. For the problem that tomatoes need to be identified in a specific environment. This study took pictures from the planting site to make the data set, A network model is built through tensorflow to classify the maturity of three kinds of tomatoes, Firstly, an improved CBAM module is added to the network to extract the information of tomato maturity and location, That is, in the channel attention module, an average pool layer after sharing multi-layer perceptron is paralleled, This method alleviates the instability problem of adding CBAM module directly to the network, Then use the Adam optimizer to update the parameters to speed up the decline of the loss function, Finally, by adjusting the learning rate and using the confusion matrix to calculate the accuracy of the verification set, the best model is obtained. The experimental results show that the loss function of the proposed network decreases steadily after training 30 epochs. A software test platform is built for testing, the accuracy is 99%, and the detection time of a single picture is 1.5s. The detection time and accuracy are better than that of alexnet network, The visualization results of Grad-CAM show that the effect of extracting target information by the network proposed in this paper is better than that of AlexNet network and CBAM module before improvement. The method proposed in this paper is suitable for melon and fruit grade classification under any background.
Keywords:
点击此处可从《科学技术与工程》浏览原始摘要信息
点击此处可从《科学技术与工程》下载免费的PDF全文
设为首页 | 免责声明 | 关于勤云 | 加入收藏

Copyright©北京勤云科技发展有限公司  京ICP备09084417号