Research question
Generated images can omit strokes, substitute characters, or mix Traditional and Simplified forms. OCR accuracy alone can miss damage to the background and nearby text.
Evaluation design
Synthetic samples contain known errors and reference text. OCR locates and classifies them. After local replacement, the prototype reassesses recognition and changes outside the mask before keeping a correction.
Current scope
The downstream pipeline runs on a CPU with synthetic samples and glyph compositing. Diffusion-based inpainting, real generated images, and generalization across fonts still require evaluation.