Oracle Bone Inscriptions (OBI), as the earliest pictographic script in China and the origin of Chinese characters, contain rich historical information. Accurate detection of OBI is crucial for their digitalization. However, challenges such as small character size and heavy noise in rubbings hinder detection performance. To address these issues, we propose a detection model named FDW-YOLO for OBI. Specifically, a Feature Focusing Diffusion Pyramid Network (FFDPN) is proposed to enhance the integration of multi-scale features. Moreover, a Dynamic Mixed Convolution Block (DMCB) is designed to reconstruct the C3K2 structure, which helps the model adaptively extract and fuse effective information under complex backgrounds. Meanwhile, to reduce the negative impact of low-quality samples on gradient propagation, we introduce WIoU v3 as the loss function. Experimental results demonstrate that FDW-YOLO achieves improvements of 2.03% in F1 score and 3.3% in mAP50 over YOLOv12. In addition, we further evaluate the proposed model by performing transfer predictions on Bronze Inscriptions and Stone Drum Inscriptions. Scalability experiments on the VOC 2007 dataset also demonstrate its robustness and adaptability. This study contributes to more efficient extraction of OBI from low-quality rubbings and provides technical support for the digitalization of ancient scripts.