Multi-scale Defective Samples Synthesis for Surface Defect Detection

Abstract

Surface defect detection has received both academic and industrial attention in recent years. In real-world applications, it is usually difficult to collect defective samples since manual labeling is time-consuming and defective samples rarely appear. In this paper, we propose a novel method for multi-scale defective sample synthesis and detection. First, a Pairs Generative Adversarial Network (PairsGAN) is proposed for generating defects and their labels. To improve the generated quality of the defective area, we design a defect discriminator in PairsGAN to focuses on distinguishing the defective area. Then, a Multi-Scale Defect Fusion (MSDF) module is presented to diversify the generated defects with various scales and styles, which fuses them into normal samples in different locations, so as to obtain naturally defective samples and corresponding labels. Finally, generated samples are used as the inputs of the semantic segmentation network for defect detection. Experimental results demonstrate that our method achieves more stable and better segmentation results comparing to recent methods.