A New Semi-Supervised Deep Learning Approach for Intelligent Defects Recognition

Abstract

Intelligent defect recognition (IDR) is one of the important technologies in production. Deep learning (DL) has drawn more and more attentions in IDR. Whereas, DL methods usually need large labelled training datasets, while the unlabeled is idle and not considered yet. In some cases, the requirement is difficult to satisfy. This is because labelling large datasets are costly, and the defect recognition might be delayed until getting enough labelled samples. To overcome this limitation, a semi-supervised DL approach for defect recognition, which uses the unlabeled samples to improve the accuracy, is introduced in this paper. This method uses a convolutional autoencoder to extract the common feature from both labelled and unlabeled samples, and only a few samples are required to finetune the network. The experimental results suggest that the proposed method achieves competitive results under limited labelled samples, and the accuracy outperforms the other approachs. Furthermore, the noise analysis also suggest this method performs robust for noisey samples.