Image Style Transfer-Based Data Augmentation for Sanitary Ceramic Defect Detection

In the task of surface defect detection on sanitary ceramics, the production environment poses limitations. There are obvious differences between the image data we collected and the image data in the actual detection scene, which leads to the performance degradation of the defect detection network in the actual working scenario. This is a well-known domain adaptation problem in the field of deep learning. A direct solution is to transform the training data to match the inference data domain. The inherent domain transformation capability of style transfer techniques renders it a potential method for addressing domain adaptation issues. Therefore, we propose a novel sample augmentation method for defects in sanitary ceramics based on a style transfer network. This method utilizes a local attention mechanism with linear computational complexity. It generates high-resolution images of sanitary ceramics with fully controllable content. Moreover, the visual features of minor defects are effectively restored through improved area-weighted loss functions. The experiments demonstrate that the application of the data augmentation method proposed in this article significantly improves the classification accuracy of all networks involved in the experiments.