Empirical Investigation of Vision Architectures for High- Cardinality Fine-Grained Ancient Chinese Ideograph Classification on OBWX

Ancient Chinese ideograph recognition is a high-cardinality fine-grained task featuring thousands of visually similar categories and a severely long-tailed data distribution. Existing studies adopt divergent dataset construction, class cardinality, and evaluation protocols, rendering their reported accuracies not directly comparable. We build a controlled empirical benchmark by evaluating 14 backbones from five architectural families with all evaluation variables held constant on the public OBWX dataset (74,047 glyph images, 6,650 categories). To disentangle the confounding effect of class frequency on single overall accuracy metrics, results are reported along four dimensions: computational cost, accuracy stratified by sample-frequency bands, 95% Wilson confidence intervals on the 12,583-sample test set, and performance degradation under controlled noise injection. EfficientNetV2-S achieves the highest accuracy (74.51%), outperforming hybrid MaxViT-T (72.31%) and hierarchical Swin Transformer (71.57%). Under the primary seed, classical CNNs lag substantially at 36.3%–36.7%, with a 32.09 percentage point (pp) performance span across architecture families. Neither parameter count nor FLOPs correlates with recognition accuracy; the largest model, ViT-B/16, underperforms by 9.07 pp despite having fourfold parameters. Accuracy decreases monotonically from head (90.32%) to tail (60.44%) categories, with tail classes accounting for 64.0% of all errors. Noise injection reverses the clean-data ranking: MaxViT-T retains 83.0% of its clean accuracy, compared with 77.4% for EfficientNetV2-S and 68.5% for Swin. Three-seed replication validates the top ranking while identifying three unstable backbones with 13.7–24.8 pp performance variations.