Application of AlexNet convolutional neural network architecture-based transfer learning for automated recognition of casting surface defects

Abstract

Automated inspection of surface defects is beneficial for casting product manufacturers in terms of inspection cost and time, which ultimately affect overall business performance. Intelligent systems that are capable of image classification are widely applied in visual inspection as a major component of modern smart manufacturing. Image classification tasks performed by Convolutional Neural Networks (CNNs) have recently shown significant performance over the conventional machine learning techniques. Particularly, AlexNet CNN architecture, which was proposed at the early stages of the development of CNN architectures, shows outstanding performance. In this paper, we investigate the application of AlexNet CNN architecture-based transfer learning for the classification of casting surface defects. We used a dataset containing casting surface defect images of a pump impeller for testing the performance. We examined four experimental schemes where the degree of the knowledge obtained from the pre-trained model is varied in each experiment. Furthermore, using a simple grid search method we explored the best overall setting for two crucial hyperparameters. Our results show that despite the simple architecture, AlexNet with transfer learning can be successfully applied for the recognition of casting surface defects of the pump impeller.