基于稠密连接卷积神经网络与混合注意力的烟叶图像分级方法  

Tobacco Image Grading Method Based on Densely Connected Convolutional Networks and Hybrid Attention

作　　者：江浩 罗瑞林 金雪松 陈载清[1] 云利军[1] JIANG Hao;LUO Rui-lin;JIN Xue-song;CHEN Zai-qing;YUN Li-jun(School of Information,Yunnan Normal University,Kunming 650500,China;Equipment Information Department,Yunnan Provincial To-bacco Company,Kunming 650218,China;Information Network Center,The Second People′s Hospital of Yuxi,Yuxi 653100,China)

机构地区：[1]云南师范大学信息学院,云南昆明650500 [2]云南省烟草烟叶公司设备信息科,云南昆明650218 [3]玉溪市第二人民医院信息网络中心,云南玉溪653100

出　　处：《软件导刊》2023年第5期184-189,共6页Software Guide

基　　金：云南省应用基础研究计划重点项目(2018FA033);中国烟草总公司云南省公司科技计划项目(2021530000242043,2022530000241026)。

摘　　要：针对烟叶实际收购过程中受环境与个人状态影响而导致分级准确率下降的问题,提出一种结合稠密连接卷积神经网络与混合注意力机制的深度学习模型。该模型通过改进残差注意力网络,在原网络注意力模块的主分支残差模块与网络输出阶段的多个残差模块上,使用稠密连接卷积模块进行替换,以增强分支特征学习性能并降低参数量,缓解梯度消失问题。同时,使用两个残差注意力网络的注意力机制模块,并添加空间注意力模块加权提取烟叶特征图在空间、通道维度的信息,以获取更全面的特征信息。通过10个等级共5 000张烟叶图像的实验表明,该网络在降低网络深度的同时,提升了检测速度与识别精度,分级正确率相较于原网络与VGGNet19分别提升8.19%、7.72%,网络参数量相较于ResNet34减少45%,训练速度提升38.11%,可证明该方法对不同等级烟叶均具有较好的识别效果和较快的识别速度,能较好地对生产中的烟叶进行分级。In view of the decline of classification accuracy caused by the influence of environment and personal status in the actual purchase process of tobacco leaves,a deep learning model combining dense connected convolutional neural network and hybrid attention mechanism is proposed.By improving the residual attention network,the model use densely connected convolution module to replace the main branch residual module in the original network attention module and multiple residual modules in the network output stage,enhance the learning performance of branch features,reduce the amount of parameters and alleviate the problem of gradient disappearance;At the same time,the two attention mechanism modules in the residual attention network are used,and by adding spatial attention mechanisms(SAM)before the two attention mechanism modules,the information of tobacco leaf feature map is weighted and extracted in the two dimensions of space and channel,so as to obtain more comprehensive feature information.Our network not only reduces the network depth,but also improves the speed and recognition accuracy.The experimental analysis of 5000 tobacco leaf images in 10 grades shows that the classification accuracy of the network is improved by 8.19% compared with the original network,7.72%compared with VGGnet19,the amount of network parameters is reduced by 45%compared with ResNet34,and the training speed is improved by 38.11%.It can be seen that this method has better recognition effect and faster recognition speed for different grades of tobacco leaves,and can be well applied to tobacco leaf classification in tobacco production.

关 键 词：烟叶分级 图像分类 混合注意力 稠密连接卷积神经网络 深度学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]