PCR-YOLO: An Improved YOLOv8 Paleographic Character Recognition Model Applying the Efficient Local Attention (ELA) Mechanism

The significance of ancient character recognition extends deeply into cultural heritage preservation, historical and linguistic research, as well as the advancement of AI technology and international collaboration. This study employs YOLOv8, a convolutional neural network-based target detection model, enhanced by the integration of the Efficient Local Attention (ELA) module into its architecture to bolster the model’s backbone in feature extraction. Consequently, the model achieves more precise and efficient extraction of ancient character content, culminating in the development of the PCR-YOLO model (Paleographic Character Recognition). Throughout the research, numerous ablation studies were conducted to ascertain the optimal placement of the ELA module, thereby defining PCR YOLO’s architecture. Comparative analyses against numerous models underscore PCR-YOLO’s superiority, with ablation results demonstrating a significant increase of 48.9% in accuracy mAP metrics, despite a modest rise in parameters and FLOPs by 0.18 M and 1.38G respectively. PCR YOLO not only topped the accuracy charts against its counterparts but also maintained a competitive edge in computational efficiency, ranking second in terms of parameter count and computational demand.