v2: Handwritten character recognition is supported

2025.9

增加手写输入，也可以直接在https://huggingface.co/spaces/raycosine/Detangutify 使用（可能需要翻墙）

Handwriting is supported. You can also use it at https://huggingface.co/spaces/raycosine/Detangutify

使用 Noto Serif Tangut （Noto 西夏宋体）作为基础字体，经过图像增强及线宽归一化，对每个字形生成多个样本，以缩小手写体与印刷体的差异。前端传入手写或上传的图像后，通过特征向量比对返回候选字形。

由于衬线印刷字体与手写体还是存在差异的，所以准确率一般。是否需要书写板正、避免连笔，不太好说；只能宽泛地说，字形复杂度与识别准确率之间没有关联，但对于某些特定偏旁部件的字形准确率会显著降低。

Using Noto Serif Tangut as the base font, multiple samples are generated for each glyph through image augmentation and stroke-width normalization to reduce the gap between handwritten and printed forms. When a handwritten or uploaded image is submitted from the frontend, candidates are returned via feature-vector matching. Since serif printed fonts differ significantly from handwriting, the overall accuracy is quite limited. Whether one needs to write neatly or avoid cursive strokes is hard to say; roughly speaking, glyph complexity does not correlate with recognition accuracy, but for certain components, the accuracy drops noticeably.

问：为什么不用手写字体训练？答曰：没有数据。