Advanced Data Augmentation Techniques for Improved Defect Detection in Industrial

Among the techniques used in industrial defect detection, reconstruction-based methods aim to reconstruct non-defective data but fail to reconstruct defective data. The anomaly score is determined based on the reconstruction error rate. However, the main challenge of these methods lies in the ineffective management of model generalization, resulting in the reconstruction of non-representative data. In this context, the emphasis is placed on defect reconstruction. To address these reconstruction issues, we employ data augmentation strategies such as Mixup, Cutmix, Cutout, and Random Erasing. These approaches mask random parts of the images, prompting the model to restore them. This improves the model’s performance in terms of reconstruction, encouraging the model to transform defective parts into non-defective ones. Our goal is to identify the technique that allows the model to reconstruct data without compromising its generalization capacity. The reconstruction is performed by a UNet architecture. The results also indicate that in evaluations with the BTAD dataset, cutmix stands out among these techniques, surpassing random erasing, cutout, and mixup by 1.4, as well as other defect detection methods by 0.8.