Automated Surface Defect Detection Based on CycleGAN Model

Abstract

Abstract Defect detection occupies an increasingly important position in the manufacturing industry, and most of approaches for the traditional defect detection are based on manual extraction of defective region traits and labeling work. This paper presents a novel defect detection approach based on Generative Adversarial Network (GAN) to automatically detect and extract defects from the target dataset. The method extends the samples using GAN model to solve the problem of insufficient samples in reality, and also provides paired samples for the second stage of defective pixel accumulation, after which the defective pixel images are output as binarized defect maps using difference accumulation and threshold segmentation. The experimental results verify that the proposed method can very accurately highlight the defects at the defect locations, and can be generated without manual labeling of defect traits.