DDCL-Net: Fine-Grained Oracle Bone Script Classification for Embodied Assistive Perception

Oracle bone inscriptions are engraved on real physical carriers such as turtle plastrons and animal bones and are commonly encountered in cultural heritage environments. Achieving stable fine-grained perception of these inscriptions in real-world physical settings remains a major challenge for embodied intelligent systems, due to severe material-induced noise, structural degradation, and subtle inter-class variations. This paper proposes a Dual-Domain Collaborative Learning Network (DDCL-Net) for fine-grained oracle bone character classification, aiming to serve as a robust visual perception module for embodied intelligent systems. DDCL-Net explicitly models structural representations across complementary frequency and spatial domains. The Phase-Aware Structure Enhancer (PASE) operates in the frequency domain to reconstruct stable character skeletons from noisy physical textures, while the Multi-Scale Gated Context Pooling (MS-GCP) module operates in the spatial domain to emulate selective visual attention and emphasize discriminative stroke regions. Furthermore, a Dual-Domain Adaptive Fusion (DDAF) module dynamically integrates structural and semantic cues, enabling robust perception under heterogeneous physical conditions. Experiments on a small-sample oracle bone character benchmark demonstrate that DDCL-Net achieves an accuracy of 92.67%, outperforming multiple mainstream visual models. From the perspective of accessible environment construction, DDCL-Net further provides a perception-level approach for inclusive cultural heritage services, enabling embodied intelligent systems to offer reliable visual assistance to users with diverse perceptual and interpretive needs. Beyond quantitative performance, the proposed framework provides a perception-level approach with potential value for embodied intelligent systems interacting with complex physical media in cultural heritage scenarios.