結合YOLO模型和特徵保留區塊方法應用於產品表面瑕疵檢測

Combining YOLO Model and FRB Method for Product Surface Defect Detection

鍾權政(Chuan-Cheng Chung)

指導教授 ： 陳永隆

共同指導教授 ： 黃馨逸(Hsin-I Huang)

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(2022年)

摘要

自動化產品表面缺陷檢測是生產線產品質量保證的關鍵項目，不僅降低人力成本、提高檢測效率也改善產品的良率，然而這項任務的關鍵取決於產品缺陷檢測的精準度。本文提出一個新的方法主要使用YOLOv5結合Feature Retaining Block(FRB)[4]，方法稱為FRB-YOLO，用於解決大部份產品在製造過程中因細微缺陷造成產品質量下降的問題，較細微的產品缺陷可能會因為與產品本身的紋路相似，而造成應該判斷屬於不良的產品卻被判斷成正常產品。本文改善因細微缺陷造成產品質量下降的問題，使用特徵保留塊FRB結合YOLOv5模型，包含兩個不同分辨率的特徵圖，並放置在YOLOv5用於萃取特徵的骨幹網路內，而神經網路會使用到池化層，池化層下採樣後會使部分缺陷特徵造成丟失，而細微缺陷的特徵在圖像內能取得的信息原本就比較少，經過池化層後細微缺陷的特徵因而丟失，使得產品缺陷檢測精準度下降，然而FRB主要的用途是用於增強被遺失的缺陷特徵訊息。此外，我們根據每個資料集的性質將不同的參數應用於不同的資料集。實驗結果表明，我們提出的 FRB-YOLO方法的mAP在DAGM上為99.01%，在Magnetic-Tile上為 72.23%，在NEU-DET上為78.75%。通過實驗，我們提出的FRB-YOLO方法的產品表面檢測精度優於YOLOv5方法的性能。

關鍵字

深度學習 ； 卷積神經網路 ； YOLO

並列摘要

The key of ensuring products’ quality in the production line is automated product surface defect inspection. It provides lower labor costs and defect rate meanwhile enhancing detection of efficiency. However, the main object depends on the accuracy of inspection. In fact, the defective products on the process of manufacturing could be recognized as normal products due to the defect are subtle, and they have similar flow mark. The issue causes the major of products qualify decreased on the manufacturing. We proposed a new method which combined YOLOv5 with feature retaining block (FRB) [4] that its name is FRB-YOLO. Our proposed method improves the issue of lowered quality by minor defects.
In our proposed FRB-YOLO method, we can extract features including incorporating the feature maps of two different recognition rates, and embedding the two feature maps in the backbone network. A neural network includes pooling layers that some defect feature of imagine is lost when the imagine downpooling. According to the information, we could be obtained from the defect features of subtle defects that it is susceptible to relatively essential scarcity. The defect features could therefore disappear that results in lowered inspection accuracy. Therefore, FRB is adopted since its main function is amplified that the defect feature information is lost. Our proposed FRB-YOLO method used the same parameters and environment on open datasets to validate our method’s effectiveness. In addition, we apply different epoch numbers to different datasets according to each dataset’s nature. Experiment results show that mAP of our proposed FRB-YOLO method is 99.01% on DAGM, 72.23% on Magnetic-Tile, and 78.75% on NEU-DET. By experiments, the accuracy of product surface inspection for our proposed FRB-YOLO method outperforms the performance of YOLOv5 method.

並列關鍵字

Deep Learning ； Convolutional Neural Networks ； YOLO

參考文獻

[1] Y. L. Chen, and H. P. Lai, "A Fuzzy Logical Controller for Traffic Load Parameter with Priority-based Rate in Wireless Multimedia Sensor Networks," Applied Soft Computing, vol. 14, pp. 594-602, Jane. 2014.

[2] D. R. Chen, "A Link- and Hop-Constrained Clustering for Multi-Hop Wireless Sensor Networks," Computer Communications, 3 2015.

[3] R. Jain, D. M. Chiu, and W. R. Hawe, "A Quantitative Measure of Fairness and Discrimination for Resource Allocation in Shared Computer System," Eastern Research Laboratory, 1984.

[4] A. Kailasand, L. Thanayankizil, and M. A. Ingram, "A Simple Cooperative Transmission Protocol for Energy-Efficient Broadcasting over Multi-Hop Wireless Networks," Journal of Communications and Networks, vol. 12, no. 8, pp. 213-220, June. 2008.

[5] A. Asadi, and V. Mancuso, "A Survey on Opportunistic Scheduling in Wireless Communications," IEEE Communications Surveys Tutorials, vol. 15, no. 4, pp. 1-18, 2013.