基于卷积神经网络和高光谱成像技术的多宝鱼新鲜度鉴别  

Freshness Identification of Turbot Based on Convolutional Neural Network and Hyperspectral Imaging Technology

作　　者：章海亮 周孝文 刘雪梅[2] 罗微 詹白勺 潘璠 ZHANG Hai-liang;ZHOU Xiao-wen;LIU Xue-mei;LUO Wei;ZHAN Bai-shao;PAN Fan(School of Electrical and Automation Engineering,East China Jiaotong University,Nanchang 330013,China;School of Civil Architecture,East China Jiaotong University,Nanchang 330013,China;Department of External Liaison,East China Jiaotong University,Nanchang 330013,China)

机构地区：[1]华东交通大学电气与自动化工程学院,江西南昌330013 [2]华东交通大学土木建筑学院,江西南昌330013 [3]华东交通大学对外联络处,江西南昌330013

出　　处：《光谱学与光谱分析》2024年第2期367-371,共5页Spectroscopy and Spectral Analysis

基　　金：国家自然科学基金项目(32260622);江西省03专项及5G项目(20212ABC03A17)资助。

摘　　要：鱼类产品新鲜度鉴别一直是重要的研究课题,相较于目前常规鱼类品质检测方法存在的成本高、检测时间长等问题,高光谱成像技术(HSI)因其无损、快速等优势得到了学者的广泛研究。卷积神经网络是深度学习中应用较为广泛的模型,表达能力强,模型效率高。因此,使用卷积神经网络(CNN)结合高光谱成像技术建立多宝鱼新鲜度鉴别模型。采集160个多宝鱼样本感兴趣区域(ROI)光谱,并根据样本不同冻融次数和冷冻时间分为5类新鲜度。以VGG11网络为基础,针对光谱数据特点对网络结构进行调整,减少全连接层数量,降低模型的复杂度,分别对比不同卷积核个数、激活函数对分类性能造成的影响,确定最佳CNN网络结构。由于高光谱数据量大同时存在的冗余信息较多,分别采用无信息变量消除算法(UVE)和随机青蛙算法(RF)对高光谱数据进行波长筛选,将波长筛选后的高光谱数据分别输入卷积神经网络(CNN)、最小二乘支持向量机(LS-SVM)、K最近邻算法(KNN)建立模型。采用无信息变量消除(UVE)提取的165个特征波长建立的UVE-CNN模型鉴别效果最佳,分类模型在测试集上的精度达到了100%。结果表明,利用卷积神经网络与高光谱成像技术相结合可实现对多宝鱼新鲜度的有效鉴别,该研究为多宝鱼新鲜度无损、准确鉴别提供新思路。Identifying fish product freshness has always been an important research topic.Compared with the problems of high cost and long detection time of the current conventional fish quality detection methods,hyperspectral imaging technology(HSI)has been obtained due to its non-destructive and rapid advantages.Convolutional neural network is a widely used model in deep learning,with strong expressive ability and high model efficiency.Therefore,a freshness discrimination model of turbot was established using a convolutional neural network(CNN)combined with hyperspectral imaging technology.First,160 regions of interest(ROI)spectra of turbot samples were collected and divided into 5 categories of freshness according to the samples'different freeze-thaw times and freezing times.Based on the VGG11 network,adjust the network structure according to the features of spectral data,reduce the number of fully connected layers,reduce model complexity,and compare the effects of different convolution kernels and activation functions on classification performance to determine the best network framework.Due to hyperspectral data and the large amount of redundant information,Uninformative variable elimination algorithm(UVE)and Random frog algorithm(RF)were used to screen the wavelength of hyperspectral data.The hyperspectral data after wavelength screening were respectively input into a convolutional neural network(CNN),least squares support vector machine(LS-SVM)and K-nearest neighbor algorithm(KNN)to establish the model.Finally,the UVE-CNN model based on the 165 feature wavelengths extracted by Uninformative variable elimination(UVE)has the best discrimination effect,and the accuracy of the classification model on the test set reaches 100%.The results showed that the combination of convolutional neural networks and hyperspectral imaging technology could be used to identify the freshness of turbot effectively.This study provides a new idea for non-destructive and accurate identification of turbot freshness.

关 键 词：卷积神经网络 多宝鱼 高光谱 深度学习 

分 类 号：O433.5[机械工程—光学工程]