Chinese Handwriting Recognition — HSK1 v5 (ResNet CNN + Embedding Verification)

A ResNet-style CNN trained on HWDB1.0 to recognise 178 Chinese characters + Unknown, with an embedding-template verification head for handwriting-practice apps.

What's new in v5

Feature	v4	v5
Mode	Classification only (which char is this?)	+ Verification (is this a correct rendering of TARGET?)
Detects wrong char	Only via low confidence	Yes — explicit cosine similarity to target template
Detects missing strokes	No (classifier picks closest)	Yes — low sim to target template even if classifier picks target
New artifact	—	`templates_v5.npz` (mean embedding per Chinese class)

What's new in v4

Feature	v3	v4
EMNIST polarity	Same as raw (bright strokes on dark bg)	Inverted to match HWDB (dark strokes on bright bg)
Brightness shortcut	Mean ratio 5.2x	Mean ratio <1.2x

Model details

Item	Value
Input	40×40 grayscale image
Classes	179 (178 Chinese characters + Unknown)
Embedding dim	512
Templates	178 (one per Chinese char)
Framework	Keras / TensorFlow
Confidence threshold (classify)	0.3
Similarity threshold (verify)	0.9
OOD training data	EMNIST Balanced (8% of training set, polarity-inverted)

Quick start — verification mode

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

model           = keras.models.load_model('chinese_hsk1_model_v5.keras')
embedding_model = keras.Model(model.input, model.layers[-2].output)

tpl_npz   = np.load('templates_v5.npz')
templates = dict(zip(tpl_npz['chars'].tolist(), tpl_npz['embeddings']))

def verify(img_gray, target_char, sim_threshold=0.65):
    x = img_gray.astype('float32') / 255.0
    x = x.reshape(1, 40, 40, 1)
    emb = embedding_model.predict(x, verbose=0)[0]
    emb = emb / (np.linalg.norm(emb) + 1e-8)
    if target_char not in templates:
        return 'invalid_target', 0.0
    sim = float(np.dot(emb, templates[target_char]))
    return ('correct' if sim >= sim_threshold else 'incomplete_or_unclear'), sim

Downloads last month: 135