因此本研究嘗試應用以卷積神經網路的方法加強金屬零件表面瑕疵檢測，目的是降低自動光學系統造成的高誤判率現況，希望能藉卷積神經網路對圖像分類領域所帶來突破性的研究成果也一併套用至檢測技術領域上，本研究以Faster R-CNN作為主架構，原始模型的整體指標(mAP)為 89.720，經過使用COCO (ResNet-50)預訓練模型，再透過轉移學習把已經訓練好的模型、參數，共享至另外一個新的模型上，進而達到優化Faster R-CNN演算法的目的，優化過後的整體指標(mAP)可達到 94.223，提高了 5% 左右。;Amidst the rhythm of rapid and large-scale production in the era of Industry 4.0, effective and precise inspection is pivotal to boost process quality. Traditionally, defect detection hinged on manual visual inspection, which proved not only time-consuming but also vulnerable to fluctuating misjudgment rates due to the inspector′s mental condition. Moreover, with an escalating complexity of metal components, the misjudgment rates of human visual inspection have an increasing tendency, leading manufacturers to adopt automated optical inspection systems in lieu of conventional manual labor. However, present automated optical inspection systems, primarily driven by image recognition, often incur high misjudgment rates in the pursuit of near-perfect detection rates, resulting in a majority of machine-selected candidate images being false defects.
This study attempts to apply convolutional neural network techniques to enhance the surface defect detection of metal components, aiming to alleviate the high misjudgment rates caused by automated optical systems. It aspires to adopt the revolutionary outcomes from the field of image classification driven by convolutional neural networks into the domain of detection technology. The Faster R-CNN model serves as the main architecture for this research. The overall Mean Average Precision (mAP) of the original model is 89.720. By employing the COCO (ResNet-50) pre-trained model and subsequent transfer learning to share the well-trained model and its parameters to a new model, the aim is to optimize the Faster R-CNN algorithm. Consequently, the overall mAP of the optimized model achieves 94.223, marking an approximate improvement of 5%.
