Application of deep learning in quality inspection of casting products

Abstract

In this paper artificial neural network is proposed as a method to classify defective and non-defective casting products in order to improve quality inspection process. Three different models of convolutional neural networks are trained and tested on dataset of submersible pump impeller images which has uneven number of image samples in each class. In order to inspect if slightly imbalanced classes have impact on result, two experiments are done. All of the models are ImageNet pre-trained networks, InceptionV3, Xception and MobileNetV2, where transfer learning method is applied with fine-tuning. Stochastic gradient descent algorithm is implemented for optimization. Obtained results of all models are presented and comparison is made.