dc.description.abstract | In the production line, in order to improve product quality, defect detection and screening is one of the important production processes. The traditional defect detection method is mainly labor-based. In addition to the high cost, the selection result will be affected by the physical condition of the personnel and environmental factors. Due to the influence of external factors, the use of computer vision for defect detection has become a more common practice in recent years.
The necessary condition for using computer vision to judge defects is that a certain amount of data is required to maintain a certain level of accuracy. However, defect data is more difficult to collect than good data, so the use of data augmentation methods is indispensable. The current common augmentation methods (such as shifting, flipping, rotating, etc.) may produce images that do not conform to the real situation, and the defect patterns are close. This method has a limited increase in the more diverse feature augmentation patterns. Therefore, this study proposes a data augmentation method for generating diverse defect data, which converts a limited number of real images into new defect data to achieve the purpose of making up for the number of defective data. In addition, in order to make the generated results under artificial control, the semantic segmentation of the labeled images is used in the generation system of this study to provide domain knowledge. However, the traditional semantic segmentation generation model is limited by a small amount of datasets, and the results are not ideal. Therefore, this study is inspired by the literature, proposes to generate a better-quality augmented image by splitting and stitching for defects. | en_US |