植基於蛋型物件擷取算法與深度學習之血蛋檢測研究

Research on Blood Egg Detection Based on Egg-Shaped Object Extraction Algorithm and Deep Learning

游政憲(Cheng-Hsien Yu)

指導教授 ： 陳民枝

共同指導教授 ： 陳同孝(Tung-Shou Chen)

國立臺中科技大學/資訊與流通學院/資訊工程系碩士班/碩士(2021年)

摘要

市售的雞蛋在進入銷售通路之前，會經過洗選蛋作業的流程將帶血的雞蛋汰除。而現有的血蛋檢測方式主要分為人工檢測與自動化設備檢測，人工檢測無法確切的管控檢測的品質，容易造成人力的浪費。而自動化的檢測設備雖然具有非常高的準確率，但其成本非常昂貴，並非大部分的蛋農能夠負擔。因此本論文提出一種血蛋檢測法，透過擷取蛋型物件及深度學習，以低成本的設備達成高準確率的血蛋檢測。本論文將雞蛋滾動的影片作為實驗資料，以模擬雞蛋在自動化產線上的轉動狀態。首先將錄製的影片切分為序列式影像，並利用本文所提出之蛋型物件擷取算法，將序列影像中的蛋體進行定位及擷取，在影像前處理的階段濾除大部分的背景雜訊。再以多種深度學習演算法及遷移式學習這項技術，進行血蛋分類模型的訓練。之後將分類模型量化後放至低成本的張量處理元件(TPU)上運行，以此提升實際使用的可能性。透過準確率的比較，選定MobileNetV2所訓練之血蛋分類模型作為本論文採用之分類模型。實驗結果顯示，本論文將血蛋分類模型放置於張量處理元件上運行，其分類準確率達到96.7%。與過往的相關研究相比，具有更高的準確率及更低的實驗成本。因此本論文之研究方法可供蛋農及這項研究領域之學者參考。

關鍵字

蛋型物件擷取算法 ； 深度學習 ； 遷移學習 ； 機械視覺 ； 血蛋檢測 ； 異常蛋檢測 ； 張量處理元件(TPU)

並列摘要

Before eggs are placed on the market, flawed eggs with blood will be eliminated through the process of egg washing. The existing blood and egg detection methods are mainly divided into manual detection and automatic equipment detection. Manual detection cannot accurately control the quality of detection, and it will also cause a waste of manpower. Although automated testing equipment has a very high accuracy rate, its cost is very expensive, which is not affordable for most egg farmers. Therefore, this paper proposed a blood egg detection method, which uses low-cost equipment to achieve high-accuracy blood egg detection through egg-shaped objects and deep learning. In this paper, the video of egg rolling is captured and used as experimental data to simulate the rotating state of eggs on an automated production line. Firstly, the recorded video is divided into sequence images, and the egg-shaped object extraction algorithm proposed in this article is used to locate and capture the egg bodies in the sequence images, and most of the background is filtered out during the image pre-processing stage Noise. Then use a variety of deep learning algorithms and transfer learning technology to train the blood egg classification model. After that, the classification model is quantified and then run on a low-cost tensor processing unit (TPU) to increase the possibility of actual use. Through the comparison of accuracy, the blood and egg classification model trained by MobileNetV2 is selected as the classification model used in this paper. The experimental results show that this paper puts the blood egg classification model on the tensor processing element and runs it, and its classification accuracy rate reaches 96.7%. Compared with previous related research, it has higher accuracy and lower experimental cost. The research method in this paper can be used as a reference for egg farmers and scholars in practice and for future research.

並列關鍵字

Egg-Shaped Object Extraction Algorithm ； Deep Learning ； Transfer Learning ； Machine Vision ； Blood Egg Detection ； Abnormal Egg Detection ； Tensor Processing Unit (TPU)

參考文獻

[1] Y. C. Liu, “Using Prism Spectrometer to Detect the Spectrum Characteristic of Blood Egg,” 中興大學物理學系所學位論文, 2014.

[2] B. L. Wang, “Improvement of Self-made Prism Spectrometer and Spectral Characteristics of Preserved Egg and Blood Spot in Eggs,” 中興大學生物物理學碩士班學位論文, 2017.

[3] G. C. Shen, “Building a PC-Based Image Inspection System to Detect the Blood Eggs with the K-Nearest Neighbor Algorithm,” 虎尾科技大學電機工程系碩士班學位論文, 2017.

[4] L. S. Chen, “Research on Duck Egg Cracked Shell Detection Based on Bagging Algorithm and Deep Learning,” 臺中科技大學資訊工程系碩士班學位論文, 2020.

[5] C. Jia, Z. H. Yu, L. Wang, Z. H. Wu, L. Wen, 基於機器視覺的禽蛋圖像分割方法研究, Journal of Northeast Agricultural University, Vol. 5, pp. 77-82, 2011.