WMAR-CB: Wavelet-Enhanced Multi-Strategy Network for Oracle-Bone Script Recognition

Applying computer vision to cultural heritage accelerates more efficient and intelligent scholarship, but artifacts pose unique challenges—notably model adaptation and robust feature representation. We observe that single-character oracle-bone images extracted from artifacts suffer from large-area noise that often blurs or even removes character structure, increasing recognition difficulty. Moreover, the scarcity of artifacts induces severe long-tail class distributions, biasing models toward high-frequency classes. To address these issues, we propose WMAR-CB, a recognition model tailored to single-character oracle-bone datasets. Our approach improves performance along two axes. Architecturally, we denoise feature maps while strengthening character-specific representations: a Wavelet Denoising Network operates across all feature-extraction stages, Asymmetric Convolution Block (ACB) enhances contour information in shallow layers, and a Multi-Head Attention (MHA) module increases focus on stroke and structure. On the training side, we introduce a sigmoid-based class-balanced loss to rebalance head–tail performance. Experiments on three benchmarks demonstrate the effectiveness of WMAR-CB: onOBC-306andOracle-AYNUwe achieve total accuracies of 92.86% and 88.36% and mean per-class accuracies of 85.39% and 83.97%, respectively; onOracle-MNISTwe reach 98.32% total accuracy.