On-Demand Multiclass Imaging for Sample Scarcity in Industrial Environments

While technology pushes towards controlling more and more complex industrial processes, data related issues are still a non-trivial problem to address. In this sense, class imbalances and scarcity of data occupy a lot of time and resources when designing a solution. In the surface defect detection problem, due to the random nature of the process, both situations are very common as well as a general decompensation between the image size and the defect size. In this work, we address a segmentation and classification problem with very few available images from every class, proposing a two-step process. First, by generating fake images using the guided-crop image augmentation method, we train for every single class a Pix2pix model in order to perform a mask-to-image translation. Once the model is trained, we also designed a automatic mask generator, to mimic the shapes of the dataset and thus create real-like images for every class using the pretrained networks. Eventually, using a context aggregation network, we use these fake images as our training set, changing every certain epochs the amount of images of every class on-demand, depending on the evolution of the individual loss term of every class. As a result, we accomplished stable and robust segmentation and classification metrics, regardless of the amount of data available for training, using the NEU Micro surface defect database.