An improved GANS model for steel plate defect detection

Abstract

Automatic steel plate defect detection is very important for it can monitor the product quality. This paper makes a study on steel plate defect detection based on machine learning. The main difficult is that there is not enough data to make powerful detection models. We propose a Generative Adversarial Networks based method to generate synthetic training image. A novel structure is designed with type related variable incorporated in Generator and a classification branch added to Discriminator. With expanded dataset, two detection algorithm, Faster R-CNN and YOLO are adopted. Various model structures, optimization methods, batch sizes and model execution time are evaluated and the influence of parameters are also analyzed. The experimental results show that the proposed novel data generation method can effectively improve the model performance.