摘 要
由於脫蠟鑄件噴砂後，表面常殘留毛面或殼模等雜質及缺陷，需再經由人工耗時的二
次檢測，本研究使用自動化光學檢測搭配深度學習進行脫蠟鑄件噴砂後之零件缺陷檢測，
目的在減少工作人員成本、並避免長時間處於高噪音及空氣品質較差之工作環境。本研究
深度學習以卷積神經網路(convolutional neural network, CNN)為主要架構，針對資料集使用
AlexNet、VGG-16、GoogLeNet及ResNet-34等四種經典卷積神經網路模型進行訓練及預測
分類有無缺陷，並將四種模型的預測結果進行綜合比較。研究結果顯示，在預測分類問題
中，除了ResNet-34之外，AlexNet、VGG-16、GoogLeNet v1在辨識上均可準確地分類出有
缺陷與無缺陷。其中，AlexNet為本研究辨識脫蠟鑄件是否有缺陷的最佳模型，其良品的預
測準確率為99.53%、不良品的預測準確率為100.00%，檢測速度約1毫秒左右。另外，本論
文設計一圖形使用者介面，此介面結合機器視覺、卷積神經網路及物件追蹤等技術，有助
於使用者操作及監控脫蠟鑄件經噴砂工法後，表面是否殘留毛面或殼模等雜質的存在。
關鍵詞：自動光學檢測、脫蠟鑄件、噴砂缺陷檢測、卷積神經網路、深度學習
ABSTRACT
In this study, the automated optical inspection (AOI) in coordination with deep learning was
employed to detect the defects such as burrs or residues of ceramic shell mold on the surface of
in sandblasted investment casting parts. The objective is to reduce the errors caused by long
hours of manual inspection and to prevent personnel from the work environments with loud noise
and poor air quality lasted long periods of time. In this study the deep learning framework wasmainly based on convolutional neural networks (CNNs). Four classic CNN models, namely,
AlexNet, VGG-16, GoogLeNet, and ResNet-34, were applied to the datasets for training,
predicting, and classifying whether there are defects and their results were compared. The results
revealed that in terms of the classification prediction, AlexNet, VGG-16, and GoogLeNet v1
could accurately determine the defects, whereas ResNet-34 could not. AlexNet was the most
accurate in detecting detects on the investment casting parts in this study; it presented a
prediction accuracy of 99.53% for good products and 100.00% for the defective ones. A
graphical user interface (GUI) including machine vision (MV), CNNs, and object tracking was
also designed and can assist users for detecting defects such as burrs surfaces or residues of shell
mold contained on the surfaces of sandblasted investment casting parts.
Keywords: Automated optical inspection, investment castings, sandblasting defect detection,
convolutional neural network, deep learning
