Oracle bone script, the earliest known Chinese writing system, holds immense cultural heritage value. Digitizing its recognition is crucial for preservation, but challenges like image degradation, noise, and cross-domain differences between handwritten and scanned data persist. To address these OBI-specific challenges, we propose Multiscale Attention and Frequency Domain Feature Fusion for Oracle Cross-Domain Recognition Networks (MAF3Net). First, an Adaptive Multi-Scale Attention Enhancement Module (AMAE) combines MultiScaleConvBlock (MSCB), AdaptiveResidualBlock (ARB), and Convolutional Block Attention Module (CBAM) to enhance feature discriminability and robustness for low-resolution oracle images. Second, a FFC-based Discriminator (FD) module employs frequency-domain analysis to explicitly capture stroke-level textures and global structural features, which are crucial for OBI. Evaluated on the Oracle-241 dataset, our method achieves 56.2% ± 0.8% accuracy on scanned data–9.1% higher than the best baseline–while maintaining 95.68% ± 0.2% accuracy on handwritten data. Although evaluated on a single dataset, this research provides a novel and tailored technical paradigm for the intelligent preservation of cultural heritage such as oracle bone inscriptions.