Feature Concentration for Supervised and Semisupervised Learning With Unbalanced Datasets in Visual Inspection

The application of deep learning to visual inspection is hampered by the scarcity of images of defective components, which are rare in modern manufacturing, and by a general lack of labeled images, because labeling is expensive. In this article, we address this by introducing feature concentration, in which features from annotated images of defective and normal components are separated in feature space by moving them towards cluster centers. We also apply feature concentration to consistency regularization in semisupervised classification, in which only a small proportion of the data is annotated. Results were compared with those from existing approaches for unbalanced and semisupervised data, using images obtained during inspection of a smartphone component. In a supervised setting, average accuracy increased by around 5%, and in a semisupervised setting, the improvement varied between 7% and 11%, depending on the supervision ratio. We also applied feature concentration to more general public datasets, where it again outperformed the other methods.