Surface Quality Augmentation for Metalworking Industry with Pix2Pix

Abstract

Image augmentation has become an important part of the data preprocessing pipeline, helping to acquire more samples by altering existing samples by cutting, shifting, etc.. For some domains, augmenting existing images is not sufficient, due to missing samples in the domains (e.g., faulty work pieces or events that occur infrequently). In such a case, new samples must be generated, since images with surface quality defects are often rare occurrence in metalworking and the amount of samples even with standard augmentation techniques does not meet requirements to train a Convolutional Neural Network (CNN) for fault detection. This paper utilizes Pix2Pix for image augmentation to generate new images with surface quality defects. The approach allows specifying the kind of defect, location, and size and transforms images by adding new defects. Furthermore, metrics to evaluate the augmented images are discussed and a recommendation of the best performing metric within the domain of metalworking is given.