[1]阮子行,黄勇,王梦,等.基于改进卷积神经网络的番茄品质分级方法[J].江苏农业学报,2023,(04):1005-1014.[doi:doi:10.3969/j.issn.1000-4440.2023.04.010]
　RUAN Zi-hang,HUANG Yong,WANG Meng,et al.Tomato quality grading method based on improved convolutional neural network[J].,2023,(04):1005-1014.[doi:doi:10.3969/j.issn.1000-4440.2023.04.010]
点击复制

基于改进卷积神经网络的番茄品质分级方法()
分享到：

江苏农业学报[ISSN:1006-6977/CN:61-1281/TN]

卷:
期数:
2023年04期
页码:
1005-1014
栏目:
农业信息工程
出版日期:
2023-08-30

文章信息/Info

Title:
Tomato quality grading method based on improved convolutional neural network
作者:
阮子行1; 黄勇1; 2; 王梦2; 史强2; 张金玲2
（1.新疆农业大学机电工程学院，新疆乌鲁木齐830052；2.新疆工程学院机电工程学院，新疆乌鲁木齐830023）
Author(s):
RUAN Zi-hang1; HUANG Yong1; 2; WANG Meng2; SHI Qiang2; ZHANG Jin-ling2
(1.College of Mechanical and Electrical Engineering， Xinjiang Agricultural University， Urumqi 830052， China;2.College of Mechanical and Electrical Engineering， Xinjiang Institute of Engineering， Urumqi 830023， China)
关键词:
番茄; 品质分级; 注意力机制; 改进卷积神经网络; AlexNet
Keywords:
tomato; quality grading; attention mechanism; improved convolutional neural network; AlexNet
分类号:
TP391.41
DOI:
doi:10.3969/j.issn.1000-4440.2023.04.010
文献标志码:
A
摘要:
为了解决番茄人工分级精度低、工作效率低等问题，基于卷积神经网络提出1种用于番茄品质分级的网络结构，并给予优化改进。设计的卷积神经网络由7个权重层（6个卷积层和1个全连接层）和4个池化层（3个最大池化层和1个全局平均池化层）构成，利用批量归一化和压缩激励模块（SE模块）进行网络结构优化。采用自采集的番茄图像数据集，通过数据增广将原1 455张图片增广至8 730张图片并进行训练和测试，用精确度、召回率、F1值（精确度和召回率的调和平均数）评估模型的各分类差异。优化后的网络模型测试精度为9657%，比未优化的网络模型测试精度提高了258个百分点。并且与传统经典网络AlexNet、MobileNet-V2、NasNet-Mobile、ShuffleNet 4种模型相比，具有收敛速度更快的优势，训练时间减少了22%～96%，测试精度提高了018～189个百分点，单张照片测试时间降低了37%～83%，计算统一设备架构（CUDA）内存占用比例也得到了一定程度的降低。优化后的网络训练过程更加稳定，模型注意力更多地集中在整个番茄上，在一定程度上降低了背景干扰，提升了算法的鲁棒性与泛化能力。通过探究3种优化算法（SGDM、Adam、RMSprop）对模型的影响，发现与Adam、RMSprop优化算法相比，SGDM优化算法的模型测试精度分别提高了201个百分点和229个百分点；去除数据集的背景后再测试单一背景对模型性能的影响，发现模型的测试精度达到9697%，相较于未去除背景的数据集，测试精度提高了040个百分点。本研究提出的卷积网络结构对简单番茄品质（分类数较少）的分级效果较好，可为番茄品质分级提供一定的理论支持。
Abstract:
In order to solve the problems of low accuracy and low work efficiency of tomato manual grading, a network structure for tomato quality grading was proposed and optimized based on convolutional neural network. The designed convolutional neural network consisted of seven weight layers (six convolutional layers and one fully connected layer) and four pooling layers (three max pooling layers and one global average pooling layer). Batch normalization and compressed excitation module (SE module) were used for network structure optimization. Using the self-collected tomato image data set, the original 1 455 images were augmented to 8 730 images by data augmentation, and then trained and tested. The classification differences of the models were evaluated by accuracy, recall rate and F1 value (the harmonic mean of accuracy and recall rate). The test accuracy of the optimized network model was 9657%, which was 258 percentage points higher than that of the unoptimized network model. And compared with the traditional classic network AlexNet, MobileNet-V2, NasNet-Mobile, ShuffleNet, it had the advantage of faster convergence speed. The training time was reduced by 22%-96%, and the test accuracy was improved by 0.18-189 percentage points, the test time of a single photo was reduced by 37%-83%, and the compute unified device architecture (CUDA) memory usage was reduced to a certain extent. The optimized network training process was more stable, and the model focused more on the tomato as a whole, which reduced background interference to a certain extent and improved the robustness and generalization ability of the algorithm. The influence of the three optimization algorithms (SGDM, Adam and RMSprop) on the model was explored. Compared with the Adam and RMSprop optimization algorithms, the model test accuracy of the SGDM optimization algorithm was improved by 201 percentage points and 229 percentage points respectively. After removing the background of the dataset, the effects of a single background on the performance of the model were tested. It was found that the test accuracy of the model reached 9697%, which was 040 percentage points higher than that of the dataset without background removal. The convolutional network structure proposed in this study has better effect on simple tomato quality grading, which provides a certain theoretical support for tomato quality grading.

参考文献/References:

[1]张超,李晓艳,汪晶晶. 绿色贸易壁垒对新疆番茄及其制品出口影响分析[J].北方园艺,2022(9):135-141.
[2]李凯,谭丹. 新疆番茄产业国际竞争力分析[J].现代商业,2018(35):65-68.
[3]缑新科,常英. 基于机器视觉的樱桃番茄在线分级系统设计[J].计算机与数字工程,2020,48(7):1798-1803.
[4]纪莹莹. 基于计算机视觉的西红柿品质分级技术研究[D].长春：吉林农业大学,2021.
[5]马高尚. 基于机器视觉的梨品质检测分级系统研究[D].秦皇岛：河北科技师范学院,2021.
[6]田有文,吴伟,卢时铅,等. 深度学习在水果品质检测与分级分类中的应用[J].食品科学,2021,42(19):260-270.
[7]BLASCO J, ALEIXOS N, MOLT E. Machine vision system for automatic quality grading of fruit[J]. Biosystems Engineering, 2003, 85(4): 415-423.
[8]SAYED G I, HASSANIEN A E, TANG M. A novel optimized convolutional neural network based on marine predators algorithm for citrus fruit quality classification[M]//LISS Proceedings of the 11th International Conference on Logistics. Singapore: Springer, 2022: 682-692.
[9]BHOLE V, KUMAR A. Mango quality grading using deep learning technique: perspectives from agriculture and food industry[C]//Proceedings of the 21st Annual Conference on Information Technology Education. USA, 2020: 180-186.
[10]何进荣,石延新,刘斌,等. 基于DXNet模型的富士苹果外部品质分级方法研究[J].农业机械学报,2021,52(7):379-385.
[11]胡发焕,董增文. 基于机器视觉的脐橙品质在线分级检测[J].湖北农业科学,2014,53(9):2160-2164.
[12]赵辉,曹宇航,岳有军,等. 基于改进DenseNet的田间杂草识别[J].农业工程学报,2021,37(18):136-142.
[13]郭小清,范涛杰,舒欣. 基于改进Multi-Scale AlexNet的番茄叶部病害图像识别[J].农业工程学报,2019,35(13):162-169.
[14]樊湘鹏,周建平,许燕,等. 基于改进卷积神经网络的复杂背景下玉米病害识别[J].农业机械学报,2021,52(3):210-217.
[15]马岽奡,唐娉,赵理君,等. 深度学习图像数据增广方法研究综述[J].中国图象图形学报,2021,26(3):487-502.
[16]唐爽. 基于深度神经网络的微表情识别[J].电子技术与软件工程,2017(3):93-95.
[17]KRIZHEVSKY A, SUTSKEVER I, HINTON G E. Imagenet classification with deep convolutional neural networks[J]. Communications of the ACM, 2017, 60(6): 84-90.
[18]刘晓龙. 面向开放集的模糊域自适应图像分类及应用[D].无锡：江南大学,2021.
[19]孙颖异,李健,时天,等. 基于改进的AlexNet卷积神经网络的植物叶片识别[J].种子,2020,39(2):77-81.
[20]叶涛涛. 基于深度卷积神经网络的X光影像分析研究[D].杭州：浙江工业大学,2020.
[21]孙俊,谭文军,毛罕平,等. 基于改进卷积神经网络的多种植物叶片病害识别[J].农业工程学报,2017,33(19):209-215.
[22]张华鹏. 基于深度学习的苹果病害识别与施药辅助决策系统研发[D].泰安：山东农业大学,2021.
[23]LIN M, CHEN Q, YAN S. Network in network[J]. arXiv preprint arXiv:1312.4400, 2013.
[24]SIMONYAN K, ZISSERMAN A. Very deep convolutional networks for large-scale image recognition[J]. arXiv, 2014, 1409：1556.
[25]IOFFE S, SZEGEDY C. Batch normalization: accelerating deep network training by reducing internal covariate shift[C]//International Machine Learning Society, IMLS. International Conference on Machine Learning. Lille, France, 2015: 448-456.
[26]HU J, SHEN L, SUN G. Squeeze-and-excitation networks[C]// Computer Vision Foundation/IEEE Computer Society. Proceedings of the IEEE conference on computer vision and pattern recognition. Salt Lake City, UT, USA, 2018: 7132-7141.
[27]王文冠,沈建冰,贾云得. 视觉注意力检测综述[J].软件学报,2019,30(2):416-439.

相似文献/References:

[1]赵丽萍,李永灿,赵统敏,等.灰霉菌毒素诱导番茄抗性突变体及相关防御酶活性[J].江苏农业学报,2016,(03):631.[doi:10.3969/j.issn.1000-4440.2016.03.023]
　ZHAO Li-ping,LI Yong-can,ZHAO Tong-min,et al.Resistant mutant induced by toxin of Botrytis cinerea in tomato and the activities of related defensive enzymes[J].,2016,(04):631.[doi:10.3969/j.issn.1000-4440.2016.03.023]
[2]李亚茹,王银磊,赵丽萍,等.番茄ty-5和 Mi-1基因多重 PCR 体系的建立[J].江苏农业学报,2016,(04):869.[doi:10.3969/j.issn.100-4440.2016.04.024]
　LI Ya-ru,WANG Yin-lei,ZHAO Li-ping,et al.Identification of ty-5 gene and Mi-1 gene by multiplex PCR in tomato[J].,2016,(04):869.[doi:10.3969/j.issn.100-4440.2016.04.024]
[3]吕敏,苏建坤,白和盛,等.桃蚜取食和机械损伤对番茄和辣椒 PAL、LOX 和 PPO 活性的诱导作用[J].江苏农业学报,2016,(06):1273.[doi:doi:10.3969/j.issn.1000-4440.2016.06.013]
　Lv?Min,SU Jian-kun,BAI He-sheng,et al.The activities of PAL, LOX and PPO in tomato and pepper plants induced by aphid herbivory and mechanical damage[J].,2016,(04):1273.[doi:doi:10.3969/j.issn.1000-4440.2016.06.013]
[4]姜静,王银磊,赵丽萍,等.番茄qRT-PCR内参基因的筛选[J].江苏农业学报,2017,(02):389.[doi:doi:10.3969/j.issn.1000-4440.2017.02.024]
　JIANG Jing,WANG Yin-lei,ZHAO Li-ping,et al.Selection of tomato reference genes for qRT-PCR[J].,2017,(04):389.[doi:doi:10.3969/j.issn.1000-4440.2017.02.024]
[5]罗伟君,唐琳,周佳丽,等.纳米锌肥对番茄果实锌含量与品质的强化[J].江苏农业学报,2016,(01):184.[doi:10.3969/j.issn.1000-4440.2016.01.028 ]
　LUO Wei-jun,TANG Lin,ZHOU Jia-li,et al.Improvement of zinc concentration and quality of tomato fruit by nano zinc fertilizer[J].,2016,(04):184.[doi:10.3969/j.issn.1000-4440.2016.01.028 ]
[6]乔俊卿,陈志谊,梁雪杰,等.枯草芽孢杆菌Bs916在番茄根部的定殖[J].江苏农业学报,2015,(06):1278.[doi:doi:10.3969/j.issn.1000-4440.2015.06.013]
　QIAO Jun-qing,CHEN Zhi-yi,LIANG Xue-jie,et al.Colonization of Bacillus subtilis Bs916 on tomato root[J].,2015,(04):1278.[doi:doi:10.3969/j.issn.1000-4440.2015.06.013]
[7]翟亚明,魏丽萍,杨倩.不同调控方式对设施盐渍化土壤特性和番茄产量及品质的影响[J].江苏农业学报,2015,(04):871.[doi:10.3969/j.issn.1000-4440.2015.04.025]
　ZHAI Ya-ming,WEI Li-ping,YANG Qian.Effects of regulatory measures on characters of greenhouse saline soil and tomato yield and quality[J].,2015,(04):871.[doi:10.3969/j.issn.1000-4440.2015.04.025]
[8]杨玛丽,赵统敏,余文贵,等.转基因 RNAi 技术在番茄研究中的应用[J].江苏农业学报,2015,(01):217.[doi:10.3969/j.issn.1000-4440.2015.01.034]
　YANG Ma-li,ZHAO Tong-min,YU Wen-gui,et al.Progress in RNA interference technique applied in tomato[J].,2015,(04):217.[doi:10.3969/j.issn.1000-4440.2015.01.034]
[9]唐玉新,曲萍,陆岱鹏,等.适合机械化移栽的番茄穴盘育苗基质配方筛选[J].江苏农业学报,2017,(06):1342.[doi:doi:10.3969/j.issn.1000-4440.2017.06.021]
　TANG Yu-xin,QU Ping,LU Dai-peng,et al.Screening of tomato plug seedling substrates proportion suitable for mechanized transplanting[J].,2017,(04):1342.[doi:doi:10.3969/j.issn.1000-4440.2017.06.021]
[10]姜静,王银磊,李亚茹,等.江苏省及其他地区番茄黄化曲叶病毒的分子鉴定及序列分析[J].江苏农业学报,2018,(01):238.[doi:doi:10.3969/j.issn.1000-4440.2018.01.035]
　JIANG Jing,WANG Yin-lei,LI Ya-ru,et al.Molecular identification and sequence analysis of tomato yellow leaf curl virus in Jiangsu province and its peripheral area[J].,2018,(04):238.[doi:doi:10.3969/j.issn.1000-4440.2018.01.035]

备注/Memo

备注/Memo:
收稿日期：2022-06-02 基金项目：新疆维吾尔自治区科协青年人才托举工程项目(RCTJ46)；新疆维吾尔自治区高校科研计划自然科学项目（XJEDU2019Y064）作者简介：阮子行(1996-)，男，安徽铜陵人，硕士研究生，研究方向为图像处理与机器视觉。(E-mail)1243450926@qq.com 通讯作者：黄勇，(E-mail)lishi182@163.com
更新日期/Last Update: 2023-09-12