Enhanced oracle bone corrosion detection using attention-guided YOLO with ghost convolution

Abstract

Abstract Oracle bone inscriptions (OBIs), as the important records of early Chinese characters, possess profound cultural and historical significance. However, These fragments are susceptible to further damage due to corrosion. Consequently, the identification and localization of corroded regions on these fragments to prevent further deterioration has emerged as a critical task. Current object detection algorithms exhibit limitations in identifying corroded regions on oracle bone fragments, which is manifested by suboptimal detection accuracy, F1 scores, and average precision (AP) values. This deficiency hinders their capability to effectively manage the complexity and diversity of corrosion patterns present on oracle bone fragments. To tackle this challenge, this study incorporates advanced attention mechanisms, including Squeeze-and-Excitation Networks (SE), Coordinate attention, and Convolutional Block Attention Module (CBAM), into the YOLOv5 detection architecture. Additionally, we introduce Ghost convolution into the backbone network to effectively retain critical feature map information while optimizing computational efficiency. Our experimental results indicate that the integration of CBAM attention and Ghost convolution within the YOLOv5 backbone markedly improves the detection accuracy of corroded regions on oracle bone fragments. Specifically, compared to the baseline YOLOv5 model, the proposed models incorporating SE, CBAM, and Ghost convolution obtain the improvements of 2.2%, 6.3%, and 8.5% in AP, respectively. This improvement in detection performance not only facilitates the identification of corroded areas but also contributes to the preservation of oracle bone heritage and the continuity of cultural knowledge.

Keywords

• Oracle bone
• Corrosion detection
• YOLO algorithm
• Attention mechanism
• Ghost convolution