Product recognition with deep learning method based on transfer learning technique

Although the share of machines and robots during production is quite high in manufacturing factories, there are sectors in which the human factor is indispensable in production. If there is man-made production at any stage of production, it is inevitable to produce faulty products. Again, human-made quality tests reduce the possibility of eliminating these faulty products before they go to the customer. In order to minimize the production of faulty products, the application of a previously trained model with faulty and correct products for the control of the products at the end of the production line will reduce the human burden working in the quality department and facilitate the detection of faulty products. In order to create a good classification model with a high success rate, this model needs to be trained with many examples. It may not always be possible to obtain a sufficient number of labeled data samples. In cases where there is not enough data to increase the model performance, it is observed in the literature with many studies that the transfer learning model increases the performance rate. In this study, a data set containing faulty and correct products obtained from a counter model was created to detect faulty products during production in a factory that produces electricity and water meters. Since the number of products in the obtained data set was low, an experimental study was carried out with the transfer learning approach. In this study, before the transfer learning technique, the data set was trained with the CNN algorithm and the success rate of the model was determined as 88%. The model created with the VGG16 algorithm, one of the transfer learning algorithms, showed a 99.5% success rate, while 99.4% with ResNet50 and 100% with Xception. As a result, it was observed that the success rate of the transfer learning algorithms in the data set used was higher than the CNN algorithm, and the Xception transfer learning algorithm, which performed the training in an average time, had the highest success rate.