Combined GANs and Classical Methods for Surface Defect Detection

Observing product quality is critical, cost-determining and time-consuming process for manufacturers. Product quality tests, on the other hand, are slow and inefficient. Human-based quality control is highly dependent on each individual controller, and traditional automated systems are both expensive and difficult to implement. With the hardware and software developments in computer vision, quality control has become fast, reliable, feasible and repeatable. In this study, we propose a deep artificial neural network-based algorithm, called DAfectNet, to detect defects on metal surfaces, uses visual data detection. During synthetic data generation, DAfectNet combines various conventional methods with a generative-adversarial network (GAN) and yields outputs that predict the class and location. While data generation with classical methods provided an improvement of 5.8% in performance as average precision, this value reached 74.95% with an increase of 8.45% using of GANs. We compared DAfectNet with the state-of-the-art methods in addition to analyzing the effects of transfer learning.