Neural font-spotting for Chinese, Japanese, and Korean

What it does

YuzuMarker.FontDetection is a PyTorch computer vision project that trains CNNs to recognize which CJK typeface produced a given text sample and to extract its typographic style. It includes a full synthetic dataset generator that composites text onto scraped Pixiv backgrounds using a third-party CJK font pack, addressing the scarcity of labeled real-world training data.

The interesting bit

The README claims this is the first-ever CJK font recognition model. The training log is endearingly honest: DeepFont failed to converge entirely, and the best ResNet-50 run hit just under 49% accuracy on synthetic data, with a footnote admitting the top score relied on a bugged augmentation.

Key highlights

• Claims to be the first publicly available CJK font recognition and style extraction model.
• Ships a complete synthetic data factory: mixes CJK characters, Japanese lyrics, Korean alphabet, and English against thousands of background images.
• Supports ResNet-18 through ResNet-50 and DeepFont backbones, built on PyTorch Lightning.
• Best reported accuracy is 48.99% on synthesized validation data using a ResNet-50 backbone.
• Dataset and an online Hugging Face demo are both available.

Caveats

• The highest reported accuracy (48.99%) was achieved with a bugged data augmentation pipeline; the fixed version scores slightly lower at 48.45%.
• All benchmarks are on synthetic data; real-world performance on photographed or scanned text is unclear.
• The Pixiv background images are randomly scraped and not redistributed, so reproducing the exact dataset requires sourcing your own backgrounds.

Verdict

Worth a look if you work with CJK typography, OCR, or need a baseline for font classification in East Asian scripts. Skip it if you need production-ready accuracy out of the box—the results are still very much lab-grade.