Dunhuang murals are an important part of human cultural heritage with extraordinary literary and historical value. However, with the influences of time and the environment, many murals have suffered large defects and require complete reconstructions of their historical information, which is difficult. Moreover, the manual restoration of murals is complicated and slow. We attempted to use deep learning techniques to restore two types of damaged murals, including central regular defects and random irregular defects. First, we created a dataset of 26,240 Dunhuang murals for model training. Second, we built an image restoration model based on U-net to automatically restore murals. We show that mapping global information to defective areas can effectively restore murals with random irregular defects. Compared to manual restoration, the proposed model significantly reduces the time and cost of restoration work. In addition, we propose a method for constructing and preserving the contour maps of mural images and present the model’s training process. Using the deep learning model, professional restorers can provide visual feedback on the restoration process to directly compare the restoration patterns of different rounds. This method implements observations of the filling-in process in the defective areas. Finally, we built a Dunhuang mural restoration system for professional restorers to restore, select, and share murals on an online platform.

The deep learning architecture YOLO (You Only Look Once) has demonstrated its superior visual detection performance in various computer vision tasks and has been widely applied in the field of automatic surface defect detection. In this paper, we propose a lightweight YOLOv8-based method for the quality inspection of car body welding spots. We developed a TA-YOLOv8 network structure which has an improved Task-Aligned (TA) head detection, designed to handle a small sample size, imbalanced positive and negative samples, and high-noise characteristics of Body-in-White welding spot data. By learning with fewer parameters, the model achieves more efficient and accurate classification. Additionally, our algorithm framework can perform anomaly segmentation and classification on our open-world raw datasets obtained from actual production environments. The experimental results show that the lightweight module improves the processing speed by an average of 2.8%, with increases in detection the mAP@50-95 and recall rate of 1.35% and 0.1226, respectively.