基於深度學習之複合型瑕疵檢測與分類

Compound Defect Inspection and Classification Based on Deep Learning

潘家聖(Chia-Sheng Pan)

指導教授 ： 陳冠宇 ； 章明

中原大學/工學院/機械工程研究所/博士(2020年)

https://doi.org/10.6840/cycu202000921

摘要

近年來整合機器視覺與平台控制的自動化檢測設備，透過引入人工智能技術，提高了產品的品質與產量，但這些自動化光學檢測設備面對較為複雜的複合型瑕疵影像檢測時，檢測誤判率仍居高不下。此外，除了檢出瑕疵外，透過了解瑕疵的類別與型態以識別加工缺失的需求亦與日俱增，本研究遂致力於開發一套具高精度與智慧化的複合型瑕疵檢測與分類系統，以深度學習技術作為瑕疵檢測與分類的基礎，透過圖形處理器大幅度提升巨量數據的運算能力，提高光學圖像檢測和分類工業產品複合型瑕疵的速度與準確性。
本研究首先藉由12288像素的線型CCD相機，以10 kHz的線速度擷取像素分辨率為3.5 μm的透明觸控面板玻璃瑕疵影像作為訓練樣本，以ZFNet網路模型經過深度學習訓練後，利用所開發的檢測與分類系統，使瑕疵分類準確率高達96%以上，且可快速及精確地分類出複合型的瑕疵種類、走向、尺寸大小與數量多寡等特徵，改善了傳統類神經網路僅能輸出單一分類結果，解決無法分類多瑕疵的缺失。
接著在印刷電路板的半成品瑕疵檢測上，採用商用自動化光學檢測設備擷取PCBA影像，並使用多種深度學習模型執行瑕疵檢測與分類的性能比較，最後建議使用YOLOv3模型來克服PCBA瑕疵的多樣性和複雜性挑戰。經由YOLOv3目標檢測框架深度學習訓練後，實測結果顯示可有效地辨識出PCBA多數量、多種類及多標籤的複合型瑕疵類別，並且精確地框選出各瑕疵的位置，其瑕疵分類準確率高達97.58%。
本研究現階段成果為面對複合型瑕疵時，已能夠快速檢測且精準分類，期盼在未來可以將自動瑕疵分類整合至光學檢查結果中，藉以判斷製造過程中的影響因素，而可對生產過程進行評估及提供改進的有效建議。

關鍵字

複合型瑕疵 ； 深度學習 ； 自動化光學檢測 ； 瑕疵檢測與分類

並列摘要

In recent years, the booming artificial intelligence technology has been gradually introduced into the defect detection system with optical images in various industrial production lines, which has improved the precision and yield of products. However, for images with compound defects, the fault detection rate is still very high. In addition, the need to identify the processing lack through defect classification is also increasing. To enhance the accuracy and intelligence of using optical images to detect and classify the compound defects on industrial products, this study compared the performance of several deep learning models in dealing with multi-defect images.
In this study, a 12288-pixel linear CCD camera was used to capture the defect images on transparent touch panel glass with a pixel resolution of 3.5 μm at a linear speed of 10 kHz. These image samples were trained by the deep learning method based on the ZFNet network model. The developed detection and classification system can achieve a defect classification accuracy of more than 96%, and can quickly and accurately classify various defect features such as defect category, inclination, size, quantity, etc.
The deep learning method is also applied to the defect detection of semi-finished printed circuit boards. Printed circuit board assembly (PCBA) images were collected by a commercial automated optical inspection equipment, and defect classification training was conducted using the YOLOv3 object detection framework. The experimental results show that the method can effectively identify multiple and compound PCBA defects in an image, and the location of each defect can be accurately framed. The classification accuracy of defects is as high as 97.58%.
In summary, this study has completed an intelligent and fast defect detection and classification system. The automatic defect classification catalogue may be integrated into the optical inspection results to propose the influencing factors in the manufacturing process. In addition, the correction of process parameters can be provided by unsupervised learning methods to improve the yield and quality of products.

並列關鍵字

Compound Defect ； Deep Learning ； Automated Optical Inspection ； Defect Inspection and Classification

參考文獻

[1] A. Mukhopadhyay, L. Murthy, M. Arora, A. Chakrabarti, I. Mukherjee, and P. Biswas, "PCB Inspection in the Context of Smart Manufacturing," in Research into Design for a Connected World: Springer, 2019, pp. 655-663.

[2] L. Xie, R. Huang, N. Gu, and Z. Cao, "A novel defect detection and identification method in optical inspection," Neural Computing and Applications, vol. 24, no. 7-8, pp. 1953-1962, 2014.

[3] Z. Chen, Y. Shen, W. Bao, P. Li, X. Wang, and Z. Ding, "Identification of surface defects on glass by parallel spectral domain optical coherence tomography," Optics express, vol. 23, no. 18, pp. 23634-23646, 2015.

[4] C. Jian, J. Gao, and Y. Ao, "Automatic surface defect detection for mobile phone screen glass based on machine vision," Applied Soft Computing, vol. 52, pp. 348-358, 2017.

[5] D. Weimer, B. Scholz-Reiter, and M. Shpitalni, "Design of deep convolutional neural network architectures for automated feature extraction in industrial inspection," CIRP Annals, vol. 65, no. 1, pp. 417-420, 2016.