集成多模型缺陷檢測影像辨識

Ensemble multi-model defect detection with image recognition

蔡昀庭(YUN-TING CAI)

指導教授 ： 曹承礎

共同指導教授 ： 陳建錦(Chien-Chin Chen)

國立臺灣大學/管理學院/資訊管理學研究所/碩士(2022年)

https://doi.org/10.6342/NTU202201805

摘要

缺陷檢測是工廠品管中重要的一環，能夠有效且精準的辨識出缺陷，為缺陷檢測的首要目標。然而過去多以人工視覺判斷或以傳統方法進行缺陷檢測，其過程耗工費時，無法維持穩定且高精準度的檢測品質，較難實際應用於工廠作業中。因此本研究提出一個貼近工廠實際需求，可應用於小型圖像數據集且能有效辨識的缺陷檢測架構，以供工廠參考並實際落地應用。
實際工廠資料集時常會有資料不平衡與影像不足的問題，本研究透過數據增強的方式擴增影像，使模型能有足夠的影像進行訓練。同時應用遷移學習技術使用Inception-v3、ResNet-50、MobileNet-v2以及EfficientNet-b0四個模型進行實驗，在有限的資源下減少模型訓練時間，並提升模型準確率。最終以集成EfficientNet-b0與MobileNet-v2的多模型組合應用可以獲得最高92.31%的整體準確率，帶來比僅使用單一模型時更好的結果。

關鍵字

缺陷檢測 ； 多模型 ； 遷移學習 ； 數據增強 ； 影像辨識

並列摘要

Defect detection plays an important role in quality control of manufacturers. The main goal of defect detection is able to recognize defects efficiently and accurately. In the past, most of manufacturers identify defect manually or use traditional methods, but it is time-consuming, costly, and hard to maintain the high accuracy of detection quality. Therefore, the previous methods are hard to be used in factory actually. This thesis proposes a defect detection framework that can identify efficiently the small datasets and meet manufacturers' need, providing manufacturers to apply actually.
In fact, datasets in factory have problems of data imbalance and insufficient image. This thesis uses data augmentation to expand image quantity, having sufficiently image to train models. Meanwhile, this thesis applies transfer learning to experiment in Inception-v3, ResNet-50, MobileNet-v2 and EfficientNet-b0 models, using this method to reduce training time and improve accuracy in limit resource. Finally, this thesis ensemble multi-model of EfficientNet-b0 and MobileNet-v2, achieving highest accuracy of 92.31%. It brings a better result than using single model.

並列關鍵字

defect detection ； multi-model ； transfer learning ； data augmentation ； image recognition

參考文獻

一、英文部分

[1]Arjovsky, M., Chintala, S., Bottou, L. (2017). Wasserstein GAN. arXiv:1701.07875.

[2]Bandara, H. E., Wijesinghe, S., Wickramasinghe, M. (2021). Identification of Degree of Defects for the Product Quality Monitoring System Using Image Edge Filtering 2021 10th International Conference on Information and Automation for Sustainability (ICIAfS).

[3]Chondronasios, A., Popov, I., Jordanov, I. (2016). Feature selection for surface defect classification of extruded aluminum profiles. The International Journal of Advanced Manufacturing Technology, 83(1), 33-41.

https://doi.org/10.1007/s00170-015-7514-3.