---
language: zh
tags:
  - image-classification
  - handwriting
  - chinese
  - keras
  - tensorflow
license: mit
---

# Chinese Handwriting Recognition – HSK1

A CNN trained on the HWDB1.0 dataset to recognise **178 Chinese characters**.

## Model details
| Item | Value |
|---|---|
| Input | 40×40 grayscale image |
| Classes | 178 Chinese characters + Unknown |
| Framework | Keras / TensorFlow |
| Confidence threshold | 0.5 (below → Unknown) |

## Files
| File | Description |
|---|---|
| `chinese_hsk1_model.keras` | Trained Keras model |
| `label_map.json` | Index-to-character mapping |
| `training_curves.png` | Accuracy & loss curves |

## Quick start
```python
import numpy as np, json
from tensorflow import keras

model     = keras.models.load_model('chinese_hsk1_model.keras')
label_map = json.load(open('label_map.json', encoding='utf-8'))

THRESHOLD = 0.5

def predict(img_40x40_gray):
    x     = img_40x40_gray.astype('float32') / 255.0
    x     = x.reshape(1, 40, 40, 1)
    probs = model.predict(x)[0]
    conf  = probs.max()
    if conf < THRESHOLD:
        return 'Unknown', conf
    return label_map[str(probs.argmax())], conf
```