TextVerifier: Seeing Beyond Recognition to Verify Chinese Text Rendering
TextVerifier catches faulty Chinese characters that recognition systems can still read, scoring 0.844 F1 on a new benchmark.
The system checks whether an image crop correctly renders a specified character, including its strokes and components. It uses structural character descriptions and training pairs showing the same character at different rendering quality. On a benchmark of 7,358 annotated characters, it outperformed OCR confidence, a general vision-language model and TextPecker. Using its scores as a per-character reward during Flow-GRPO post-training also reduced rendering errors.
HF Daily Papers' note
The system checks whether an image crop correctly renders a specified character, including its strokes and components. It uses structural character descriptions and training pairs showing the same character at different rendering quality. On a benchmark of 7,358 annotated characters, it outperformed OCR confidence, a general vision-language model and TextPecker. Using its scores as a per-character reward during Flow-GRPO post-training also reduced rendering errors.
HF Daily Papers' note
score 5