A simulation-based few samples learning method for surface defect segmentation

Abstract

In industrial production, it is difficult to obtain a well-trained surface detection algorithm since the real defect samples are lacking. In this paper, we propose a surface defect segmentation method based on defect sample simulation, which only needs few defect training samples. The entire method includes two modules: a local defect simulation algorithm and a residual-restored-based segmentation algorithm. In order to ensure both structural and local texture consistency of the simulated defects, we design a two-stage simulation algorithm based on generation adversarial net and neural style transfer. The simulation method requires one single defect reference sample for training, and can generate the same type of defect in the specified area. The segmentation algorithm, trained with the simulated images and reference samples, can restore the defect area and yield the predicted label from the residual image. We carry out experiments on the button, road crack, and silicon steel strip datasets. The results show that the proposed method can remarkably improve the defect segmentation accuracy, attaining F1 score of 0.82.