A novel deep learning motivated data augmentation system based on defect segmentation requirements

Deep learning methods, especially convolutional neural networks (CNNs), are widely used for industrial surface defect segmentation due to their excellent performance on visual inspection tasks. However, the problems of overfitting and low generalizability affect the performance of CNN-based surface defect segmentation models. Therefore, data augmentation is necessary to reduce overfitting and improve generalization. However, existing data augmentation methods are random and independent of the downstream defect detection tasks. Thus, we propose a simple plug-and-play data augmentation method based on the requirements of the CNN defect segmentation task. We first pretrain a defect segmentation model on a training set and obtain confidence maps. Then, we occlude high-confidence regions based on the data augmentation module to balance the attention paid by the model to high- and low-confidence regions. Finally, to prevent overfitting, we periodically update the confidence map. We conduct sufficient experiments on the Kolektor surface defect (KSD) metal surface dataset and an optoelectronic chip dataset. The proposed method can make full use of the information contained in the existing datasets to improve the accuracy of the segmentation model (intersection-over-union (IOU) 6.62% improvement over baseline on the KSD dataset and 3.20% on the optoelectronic chip dataset). Moreover, the proposed method is highly applicable to various mainstream defect segmentation methods.