A Deep Learning Based End-to-end Surface Defect Detection Method for Industrial Scenes

Abstract

Due to the large amount of data required for deep learning, annotation is costly and defective sample data is difficult to obtain in industrial production. To address these issues, a two stage network is proposed in this paper that requires only a small number of pixel level labels, introducing various optimization strategies to reduce the need for high precision annotation. The method is evaluated using the DAGM dataset as well as an industrial field acquisition dataset created. Experimental results show that the method outperforms some classical methods under mixed supervision, with significantly lower annotation costs. The improved network using all categorical labels and 25% pixel labels achieved 93.9% mAccuracy on the DAGM dataset and 85.63% mAccuracy on the self-constructed dataset.