Study of Data Enhancement Techniques for Steel Surface Defect Detection

The steel surface defect detection is always worth exploring; for this, Convolutional Networks for Biomedical Image Segmentation (U-Net) is one of the commonly used techniques for solving such problems, which is often used with Residual Network (ResNet) or other encoders. However, apart from the main technique structure, data augment methods for segmentation problems have not been widely and deeply studied, which is the focus of this paper. The purpose of this research is to investigate the impact of different kinds of data augmentation methods on the model training effect. Specifically, this study mainly trained the model by applying different kinds of data augmentation methods separately and comparing them through the set metrics: Dice coefficient and Intersection over Union (IoU). Different kinds of data augmentation methods were grouped according to their functions. After obtaining the eventual data of each group, the good-performing ones were selected for further combining or adjusting the parameters. This study is conducted on Severstal: Steel Defect Detection from the Kaggle platform. The final results show that data augment methods do not necessarily lead to a huge improvement in the metrics with the used data set and the framework. The research also analyzes the potential reasons for the result. Moreover, the prevention of overfitting by these methods is still not negligible. The study has value for the use of data augment methods, especially in the treatment of segmentation problems, and has some implications for more in-depth research.