Dual-Path GAN: A Method for Enhancing Small-scale Defect Detection on Metal Images

Automated surface inspection (ASI) is an important research content in computer vision. In recent years, with the application of deep learning models represented by convolutional neural networks (CNN) in computer vision, surface defect detection based on computer vision has made impressive progress. However, compared with a few or dozens of pictures in real industrial scenes, traditional deep learning methods require a large amount of annotation data for training, so it is quite difficult to adapt to the complex industrial scenarios with varying surface profiles, lighting conditions, imaging angles and environments. Despite these difficulties, there is still a wide gap in the performance between the detection of small-scale and large-scale objects. Traditional Generative Adversarial Nets (GANs) based augmentation method can only be used for classification networks or unsupervised learning, and if applied to detection networks, they need to be labeled, which requires much labor and time. In order to solve these problems, a metal surface small-scale defect images augmentation method is proposed. Our method contains two parts: generation part and augmentation part. In the generation part, two pairs of generators and discriminators are used to generate the defect areas and the background areas of the image, which can not only generate more realistic defect images than others but also save the training process for surface images. In the augmentation part, the real defect images and the generated defect images are concatenated before copy-pasting them on background images to generate annotated dataset for training. We conduct experiments on our metal surface dataset. The experimental results show that our method can generate high-quality defect samples and background samples, which greatly enriches the original dataset. We evaluate different augmentation strategies, and ultimately, we achieve 6.2% improvement on baseline and 4.4% on copy-pasting method.