王博 楊洪遙 陸逢貴 陳子?xùn)| 曹振霞 劉登勇
摘 要:以重組牛肉為研究對(duì)象,基于機(jī)器視覺(jué)技術(shù)構(gòu)建3 種深度殘差網(wǎng)絡(luò)(deep residual network,ResNet)模型(ResNet-50、ResNet-101、ResNet-152)用于識(shí)別重組牛肉,同時(shí)應(yīng)用VGG-16視覺(jué)幾何群網(wǎng)絡(luò)模型、支持向量機(jī)模型以及LeNet-5卷積神經(jīng)網(wǎng)絡(luò)模型,比較分析ResNet模型的識(shí)別準(zhǔn)確率和響應(yīng)時(shí)間。采集并經(jīng)過(guò)圖像預(yù)處理后共得到6 168 張樣品圖像作為實(shí)驗(yàn)樣本,隨機(jī)選取其中的4 936 張作為訓(xùn)練集,剩余1 232 張作為測(cè)試集。結(jié)果表明:3 種ResNet模型(ResNet-50、ResNet-101、ResNet-152)識(shí)別速率較快,準(zhǔn)確率高,均可以有效識(shí)別重組牛肉,且卷積層越多,準(zhǔn)確率越高,其中ResNet-50模型識(shí)別準(zhǔn)確率達(dá)到較高水平,且測(cè)試時(shí)間僅需0.45 s,能夠準(zhǔn)確、快速地識(shí)別重組牛肉。
關(guān)鍵詞:重組牛肉;識(shí)別;卷積神經(jīng)網(wǎng)絡(luò);深度殘差網(wǎng)絡(luò)
Abstract: Three deep residual network (ResNet) models (ResNet-50, ResNet-101 and ResNet-152) to quickly identify restructured meat were built based on machine vision technology, and they were comparatively analyzed for recognition accuracy and response time applying visual geometry group network (VGG-16) model, support vector machine (SVM) model and LeNet-5 convolution neural network model. Images of restructured beef steak samples were collected and preprocessed. As a result, a total of 6 168 images were obtained for this research, 4 936 of which were randomly selected as the training group, and the remaining 1 232 were used as the test group. The results showed that all the three ResNet models could fast and accurately identify restructured beef steak. With more convolution layers, the accuracy was higher. The ResNet-50 model exhibited higher recognition accuracy with testing time of only 0.45 s and it was a better one to accurately and quickly identify recombined ground beef.
Keywords: restructured beef; recognition; convolutional neural network; deep residual network
DOI:10.7506/rlyj1001-8123-20200521-133
中圖分類號(hào):TS251.52? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ? ?文獻(xiàn)標(biāo)志碼:A 文章編號(hào):1001-8123(2020)07-0013-05
引文格式:
王博, 楊洪遙, 陸逢貴, 等. 重組牛肉圖像識(shí)別模型的比較研究[J]. 肉類研究, 2020, 34(7): 13-17. DOI:10.7506/rlyj1001-8123-20200521-133.? ? http://www.rlyj.net.cn
WANG Bo, YANG Hongyao, LU Fenggui, et al. Comparative study on image recognition models for restructured beef[J]. Meat Research, 2020, 34(7): 13-17. DOI:10.7506/rlyj1001-8123-20200521-133.? ? http://www.rlyj.net.cn
重組肉是指通過(guò)機(jī)械和添加輔料使肌肉中的肌原纖維蛋白析出,然后利用添加劑的黏合作用使肉糜、肉顆?;蛉鈮K重新組合,經(jīng)冷凍后直接出售或經(jīng)預(yù)熱處理保留和完善其組織結(jié)構(gòu)的肉。在我國(guó)重組肉種類較多,如重組牛排、培根等,產(chǎn)量也極大。碎肉塊經(jīng)過(guò)黏合劑及香精、香料處理后,具有口感豐富、肉質(zhì)鮮嫩等特點(diǎn)[1]。目前,由于國(guó)內(nèi)市場(chǎng)上存在重組肉及其制品生產(chǎn)標(biāo)準(zhǔn)不夠完善、相關(guān)監(jiān)管部門的管理力度欠缺等問(wèn)題,導(dǎo)致一些不法生產(chǎn)者在實(shí)際生產(chǎn)中為牟取利益,用劣質(zhì)碎肉原料取代優(yōu)質(zhì)原料,嚴(yán)重者直接以低價(jià)肉取代高價(jià)肉,如以碎鴨肉添加香精、香料仿制重組牛排,不僅損害了消費(fèi)者利益,而且對(duì)市場(chǎng)的良性發(fā)展造成極大的安全隱患[2-3]。
因此建立快速、準(zhǔn)確識(shí)別重組碎肉的方法,對(duì)規(guī)范市場(chǎng)秩序和指導(dǎo)消費(fèi)者準(zhǔn)確選購(gòu)均具有重要意義。