Chinese Handwriting Recognition — HSK1 v3 (ResNet CNN)

A ResNet-style CNN trained on HWDB1.0 to recognise 178 Chinese characters + Unknown.

What's new in v3

Feature	v2	v3
Architecture	Plain CNN	ResNet-style residual blocks
Loss	Cross-entropy	Cross-entropy + label smoothing (0.1)
Optimizer	Adam	AdamW (weight decay 1e-4)
Augmentation	Rotation, shift, zoom	+ Shear, stronger zoom
Temperature scaling	Yes (buggy)	Removed (not needed)
Config	Scattered	Central CFG dict

Model details

Item	Value
Input	40×40 grayscale image
Classes	179 (178 Chinese characters + Unknown)
Framework	Keras / TensorFlow
Confidence threshold	0.3
OOD training data	EMNIST Balanced (8% of training set)

Quick start

import numpy as np, json
import tensorflow as tf
from tensorflow import keras

model     = keras.models.load_model('chinese_hsk1_model_v3.keras')
label_map = json.load(open('label_map_v3.json', encoding='utf-8'))
cfg       = json.load(open('config_v3.json'))
THRESHOLD = cfg['threshold']

def predict(img_gray):
    x     = img_gray.astype('float32') / 255.0
    x     = x.reshape(1, cfg['img_size'], cfg['img_size'], 1)
    probs = model.predict(x)[0]
    conf  = probs.max()
    idx   = probs.argmax()
    char  = label_map[str(idx)]
    if char == 'Unknown' or conf < THRESHOLD:
        return 'Unknown', conf
    return char, conf

Downloads last month: 57