整合WGAN-GP及YOLOv5於不平衡鋼帶金屬表面資料集之瑕疵檢測
|
| 摘要
在鋼帶生產環境中，機台設備以及環境因素的影響，導致鋼帶產生表面缺陷，並且對於鋼鐵業來說，表面缺陷是對產品質量的最大威脅，使用有效且即時的瑕疵檢測技術為生產高質量產品的關鍵。然而，在工業領域實際生產中收集到的瑕疵樣本數量有限，各瑕疵類別數量也經常處在一個不平衡的狀態下，致使訓練物件檢測模型效果不佳。因此本研究使用鋼帶表面瑕疵資料集為例，利用基於梯度懲罰的Wasserstein生成對抗網路用於生成瑕疵樣本數量，已達資料平衡之目的，並透過訓練YOLOv5模型進行瑕疵檢測，最終評估驗證是否能有效的使用在實際生產環境瑕疵樣本在數量不平衡狀態下的瑕疵檢測。 關鍵字：瑕疵檢測、鋼帶瑕疵、生成對抗網路、WGAN-GP、YOLOv5。 ABSTRACT In the steel belt production environment, the influence of equipment and environmental factors leads to surface defects in steel belts, and for the steel industry, surface defects are the biggest threat to product quality, and the use of effective and real-time defect detection technology is the key to producing high-quality products. However, the number of defect samples collected in the actual production of the industrial field is limited, and the number of defect categories is often in an unbalanced state, resulting in the poor effect of the training object detection model. Therefore, this study uses the steel belt surface defect dataset as an example, and uses the Wasserstein generative adversarial network based on gradient penalty to generate the number of defect samples, which has achieved the purpose of data balancing, and trains the YOLOv5 model for defect detection, and finally evaluates and verifies whether the defect detection in the actual production environment can be effectively used in the unbalanced number of defect samples. KEYWORDS: Defect Detection; Steel Belt Defect; Generative Adversarial Network; WGAN-GP; YOLOv5. |