詳細書目資料

30
0
0
0

應用深度學習的視覺辨識對塑膠製品缺陷檢測 = Visual Recognition with Deep Learning for Defect Detection of Plastic Products / 李應賢.

館藏資訊

摘要註

塑膠製品的質量抽檢檢測是工業生産重要的環節，傳統人工檢測方法 存在效率低、主觀性極强且費時，對現代化生産環節顯得落伍。本研 究提出應用深度學習的視覺辨識，來對塑膠產品在產線上的抽查檢驗 的缺陷高效能辨識方法。本研究使用 YOLOv8n 實時性和高效性，並通過改進的特徵提取辨識，對檢測精度與速度的效率做提升，研究方 面先構建了一種常見塑膠產品缺陷，定位柱斷裂是常見缺陷的數據 集， 針對其進行標注和預處理。使用深度學習技術， 預訓練的YOLOv8n 模型對該數據集進行訓練，得到塑膠定位柱斷裂檢測的專用模型。因同時評估分類準確性與框選位置的精確度。相較於傳統的 分類準確率，mAP 更能全面反映目標檢測任務的辨識與定位能力。人工檢查的準確率約在 80% 左右，自動化抽檢中實際應用 85%為實用等級，所以本研究設定模型在測試數據集上取得了 mAP（mean Average Precision）設定超過 85%的檢測精度，且能毫秒級內完成單張圖片的檢測。. Quality inspection through sampling is a critical step in the industrial manufacturing of plastic products. Traditional manual inspection methods are inefficient, labor-intensive, and highly subjective, making them no longer suitable for modern production lines. To address these limitations, this study proposes an efficient approach for defective detection using deep learning-based visual recognition. Specifically, the YOLOv8n model is employed due to its real-time performance and computational efficiency. The model’s feature extraction and recognition capabilities have been enhanced to improve both detection accuracy and inference speed. A custom dataset was constructed focusing on a common plastic product defect— fractures in positioning pillars—which were annotated and preprocessed for training. Leveraging transfer learning, the YOLOv8n model was fine-tuned on this dataset to develop a dedicated defect detection model for fractured positioning pillars. To evaluate both classification accuracy and object localization precision, this study utilizes Mean Average Precision (mAP) as its primary performance metric. Compared to traditional classification accuracy, mAP provides a more comprehensive evaluation of a model’s object detection capabilities. While manual inspection typically yields accuracy of around 80%, industrial applications often require at least 85% accuracy for practical deployment. Accordingly, this research sets a performance target of achieving a mean Average Precision (mAP) of greater than 85% on the test dataset. The final model meets this requirement and demonstrates the ability to detect defects in individual images within milliseconds, supporting the feasibility of deploying real-time automated quality inspection in manufacturing environments..

回到最上