基于贝叶斯超参数优化的鲜香菇机器视觉图像分级识别  

Classification and recognition of fresh Lentinus edodes with machine vision image processed by Bayesian hyperparametric optimization technology

作　　者：张瑞青[1,2] 贺智斌 陈文杰 李张威 郝建军 ZHANG Ruiqing;HE Zhibin;CHEN Wenjie;LI Zhangwei;HAO Jianjun(College of Mechanical and Electrical Engineering,Hebei Agricultural University,Baoding 071001,China;Hebei Province Smart Agricultural Equipment Technology Innovation Center,Baoding 071001,China;School of Life,Langfang Normal University,Langfang 065000,China)

机构地区：[1]河北农业大学机电工程学院,河北保定071001 [2]河北省智慧农业装备技术创新中心,河北保定071001 [3]廊坊师范学院生命学院,河北廊坊065000

出　　处：《河北农业大学学报》2024年第5期116-123,共8页Journal of Hebei Agricultural University

基　　金：河北省现代农业产业技术体系创新团队项目(HBCT2018050201).

摘　　要：针对鲜香菇分级机械化程度低,精度不高等问题,本文提出1种基于贝叶斯超参数优化技术的鲜香菇机器视觉图像识别方法。利用摄像头拍摄鲜香菇图像,按人工分级标准对采样图像进行正反面标记分级,获取并标记了5级的鲜香菇图像,利用仿射变换和对比度变换的方法对获取的数据集进行扩充,建立各等级鲜香菇图像数据集;基于深度卷积神经网络,对3种预训练网络模型(AlexNet、GoogLeNet、ResNet-18)分别进行迁移学习,3种模型分别记为XGu_Ale、XGu_Goo和XGu_Res-18;使用贝叶斯优化算法对3种模型的香菇正反面数据集进行超参数优化,并分析了各个网络模型的测试结果。分析可知鲜香菇正面图像等级模型以Z-XGu_Res-18模型的识别准确率最高,鲜香菇反面图像等级模型以F-XGu_Res-18模型的识别准确率最高,准确率分别为98.73%和99.15%,选择以上2个模型可满足鲜香菇的分级要求,对正反面识别结果进行加权组合得到鲜香菇分级识别的最终等级。Aiming at the problems of low mechanization and low accuracy of fresh mushroom classification,this paper proposed a machine vision image recognition method of fresh mushroom based on Bayesian hyperparametric optimization technology.The camera is used to shoot fresh mushroom images,and the sampled images were marked and classified according to the manual classification standard.Five levels of fresh mushroom images were obtained and marked.The obtained data set was expanded by affine transformation and contrast transformation,after which the fresh mushroom image data set of each level was established;Based on the deep convolution neural network,three pre training network models(AlexNet、GoogLeNet、ResNet-18)were transferred and learned respectively,and the three models were recorded as XGu_Ale,XGu_Goo and XGu_Res-18;Bayesian optimization algorithm was used to optimize the super parameters of mushroom data sets of three models,and the test results of each network model were analyzed.The analysis showed that the front image level model of fresh Lentinus edodes was based on Z-XGu_Res-18 model had the highest recognition accuracy for reverse image hierarchy model of fresh Lentinus edodes is based on F-XGu_Res-18 model had the highest recognition accuracy,which was 98.73%and 99.15%,respectively.The above two models can meet the classification requirements of fresh mushrooms.The weighted combination of the positive and negative recognition results can obtain the final classification of fresh mushrooms.

关 键 词：图像识别 贝叶斯超参数优化 鲜香菇分级 迁移学习 

分 类 号：S565.2[农业科学—作物学] TP391.4[自动化与计算机技术—计算机应用技术]