Detection of Defects in Work-in-Process Products Using Deep Learning: A Case of a Water Tap Production Line

This paper presents the development of a deep learning model designed to identify two classes of object images, the work-in-process of a kind of tap. There were the good parts of316 images and the defective parts of320 images. Both classes of images were processed using the oversampling technique for data augmentation before transformation. Then, the processed data were applied to training three transfer learning models, consisting of ResN et50, MobileN et, and Xception. The results show an accuracy and F1-score of 100
^%
for the ResN e50 model, while the augmentation can help improve the performance of the MobileNet and Xception models. Particularly, when evaluated with the test set, the accuracy can be improved from 98.96% to 99.67% for the MobileNet model, while it can be increased from 95.83% to 97.01 % for the Xception model. The accuracy rates were slightly improved, about 0.72% and 1.23% for MobileNet and Xception, respectively. This means that the augmentation approach can be used to improve the performance of transfer learning models. For the major contributions of this study, it could be extended to other types of taps, and then ResNet50, which is the best deep learning model in this study, can be applied for the development of a new station in a production line, for detection and rejection defects automatically.