Defect Detection Using Combined Deep Autoencoder and Classifier for Small Sample Size

Defect detection is a crucial step in the process of manufacturing computer keyboards. Light leakage is a major class of defects. The keyboards with light leaking are not considered as quality products. Currently, camera images are used for light leakage detection. One major problem of the conventional computer vision-based detection algorithm is false positive which misclassifies the dust as defects. In this paper, we propose a novel algorithm using deep neural networks combining autoencoder with fully connected network (FCN) to distinguish the light leakage defect from mere dust. The proposed deep learning network architecture can improve the generalization performance by imposing much more constraints from autoencoder, and be suitable for applications with small training data sample sizes. The experimental results show that the proposed deep learning method can significantly reduce the defect type II error from 6.27% to 2.37% while the dust detection accuracy is comparable.