Data augmentation strategy for generating realistic samples on defect segmentation task

Massimo Martini, Riccardo Rosati, Luca Romeo, Adriano Mancini

Abstract: In the anomaly and defect detection tasks, the number of negative samples greatly exceeds the number of defective samples. As a result, a high-class imbalance exists among different classes in the detection task. In our work, we introduce a data-level solution for improving the generalization performance of the semantic segmentation of surface defects based on a data augmentation (DA) strategy. In particular, our DA approach comprised a generative stage to simulate synthetic defects and a validation stage to validate the synthetic image as close as possible to the real one. A Siamese network fully validates our synthetic samples to select only synthetic defects as close to the real ones. We demonstrated the effectiveness of our approach in a real-use case scenario to baseline DA approaches. Our DA approach allows balancing the minority classes while improving the overall generalization performance for semantic segmentation for defect detection.