人工注脂牛肉之自動化檢測

Automated inspection of artificial marbling beefs

謝伊婷(YI-TING HSIEH)

指導教授 ： 林宏達

朝陽科技大學/理工學院/工業工程與管理系/碩士(2020年)

摘要

牛肉注脂技術的出現是為了使原來等級較低的肉，提升食用價值而發展出的技術，而在經過脂肪注射後的人工注脂牛肉擁有了與和牛相同含有密集脂肪分布的外觀特徵。由於和牛價格昂貴而注脂牛肉成本低，因此市面上出現了一種「和牛級」的人工注脂牛肉，許多業者將注脂牛肉打著和牛的名號，以高單價的價格出售謀取暴利，而消費者卻以高價錢買到低品質的牛肉，且可能產生食品安全疑慮。因此本研究針對此狀況提出一套注脂牛肉檢測系統，期望能夠提供消費者在選購牛肉時，利用手持行動裝置擷取牛肉影像，對牛肉表面紋路與顏色進行辨別，資料經傳輸至伺服器端分析後，即可獲得影像分類結果，可即時得知該待測影像是否為人工注脂牛肉，進而保障消費者的權益以減少食品詐欺的問題。

本研究在取得手持裝置所擷取之牛肉影像後，使用 ROI (Region Of Interest)遮罩排除影像之背景與干擾物後進行影像分格，接著分別提取影像分格區塊之局部二值模式(Local Binary Patterns, LBP)紋路特徵及 RGB 色彩特徵後，使用支援向量機(Support Vector Machine, SVM)模式進行分格區塊影像分類，最後則採多數決方式判別該待測影像之類別。本研究使用 360張訓練影像及 180 張測試影像進行實驗，實驗結果顯示可以有效地辨別出三種不同牛肉類別，其注脂牛肉檢出率(1-)為 95.00%、注脂牛肉誤判率()為 1.67%、正確分類率(CR)為 93.89%及 F1-指標(F1-Score)為 95.80%。

關鍵字

食品詐欺 ； 人工注脂牛肉檢測 ； 電腦視覺 ； 局部二值模式 ； 紋路特徵 ； 機器學習

並列摘要

In order to make the original lower-price beef into higher-price beef and increase food value, the technology of injecting fat in beef was invented.
The fat-injected beef is called “artificial marbling beef” and it has the same appearance characteristics with dense fat distribution as those of Wagyu beef. Since Wagyu beef is expensive and the artificial marbling beef is cheap, the “Wagyu-grade” artificial marbling beef becomes common in food fraud. Many sellers sell the artificial marbling beef under the name of Wagyu beef for higher price to earn substantial profit. This leads to consumer using high price to purchase not only low-grade beef but also with food safety concerns. Therefore, this study proposes an automated inspection system of artificial marbling beefs and expect to solve the food fraud problems for protecting consumer’s right.

In this study, after obtaining the beef images by using handheld devices, we use ROI (Region of Interest) mask to get rid of the background and other interference items from the original image. And then, we grid the image into many equal-sized blocks. After the above pretreatments, we can extract the texture features LBP (Local Binary Patterns) and RGB color vectors from each gridding block. We apply these feature vectors as input to classify these gridding blocks into three beef categories by SVM (Support Vector Machine) model. Finally, we take the majority to determine the beef category of each image. We use 360 training images and 180 testing images to carry out the experiments. The experimental results show the proposed system can effectively achieves 95.00% artificial marbling beef detection rate, 1.67% non-artificial marbling beef false alarm rate, 93.89% correct classification rate (CR) and 95.80% F1-score.

並列關鍵字

Food fraud ； Inspection of artificial marbling beef ； Computer vision ； Local Binary Patterns ； Texture feature ； Machine learning

參考文獻

[1] Abuzneid, M. A., and Mahmood, A., “Enhanced human face recognition using LBPH descriptor, multi-KNN, and back-propagation neural network,” IEEE Access, Vol. 6, pp. 20641-20651, (2018).

[2] Akram, M. W., Li, G., Jin, Y., Chen, X., Zhu, C., Zhao, X., Khaliq, A., Faheem, M., and Ahmad, A., “CNN based automatic detection of photovoltaic cell defects in electroluminescence images,” Energy, Vol. 189, 116319, (2019).

[3] Arsalane, A., El Barbri, N., Tabyaoui, A., Klilou, A., Rhofir, K., and Halimi, A. “An embedded system based on DSP platform and PCA-SVM algorithms for rapid beef meat freshness prediction and identification,” Computers and Electronics in Agriculture, Vol. 152, pp. 385-392, (2018).

[4] Backes, A. R., and Junior, J. J. d. M. S., “LBP maps for improving fractal based texture classification,” Neurocomputing, Vol. 266, pp. 1-7, (2017).

[5] Backes, A. R., and Junior, J. J. d. M. S., “LBP maps for improving fractal based texture classification,” Neurocomputing, Vol. 266, pp. 1-7, (2017).