Japanese Character Recognition CNN

A Convolutional Neural Network (CNN) for classifying individual Japanese characters from 64x64 grayscale images. Three model versions are available.

v3 — PyTorch/ONNX (recommended)

KanjiDNN_v3.onnx — a compact 4-block CNN optimized for finger-drawn input on touchscreens.

Property	Value
Framework	PyTorch, exported to ONNX
Classes	3,181 (2,140 Joyo kanji + 859 JIS Level 1 kanji + 181 kana + 1 REJECT)
Input	64x64 grayscale, binarized, ink=1.0 / background=0.0
Parameters	~2.4M
File size	11 MB
Training data	KanjiVG SVG strokes with finger-drawing augmentation + ETL9G handwriting samples

Architecture

4x ConvBlock(Conv3x3 → BN → ReLU → Conv3x3 → BN → ReLU → MaxPool2x2 → Dropout)
  channels: 32 → 64 → 128 → 256
AdaptiveAvgPool → Linear(256,512) → BN → ReLU → Dropout(0.4) → Linear(512, num_classes)

Preprocessing

Images must be preprocessed identically to training (see KanjiDNN_v3_preprocessing_config.json):

Binarize at threshold 200
Crop to ink bounding box
Add margin (10% of the larger dimension)
Resize to 64x64 (preserve aspect ratio, center on canvas)
Normalize: (255 - pixel) / 255 (ink=1.0, background=0.0)

Rejection

Dual-layer rejection (see KanjiDNN_v3_inference_config.json):

Explicit REJECT class (index 3180) trained on romaji, digits, punctuation, scribbles, and blanks
Confidence threshold: force reject when top softmax probability < 0.5

Usage (ONNX Runtime)

import onnxruntime as ort
import numpy as np
import json

session = ort.InferenceSession("KanjiDNN_v3.onnx")
classes = json.load(open("KanjiDNN_v3_classes.json"))

# image: preprocessed 64x64 float32 array, shape (1, 1, 64, 64)
image = np.expand_dims(np.expand_dims(preprocessed, 0), 0).astype(np.float32)
logits = session.run(None, {"input": image})[0]

probs = np.exp(logits) / np.exp(logits).sum()
top_idx = np.argmax(probs)
print(f"Prediction: {classes[top_idx]} ({probs[0][top_idx]:.1%})")

Why v3?

Built to replace v1/v2, whose precision suffers on real finger input — their inference pipeline skips bounding-box centering and normalization. v3 uses KanjiVG strokes augmented to approximate finger-drawing (thick, imprecise strokes, no pressure taper) plus ETL9G handwriting samples, with a rigorous preprocessing pipeline that must be replicated at inference time.

Source code

Training pipeline: github.com/gaiseras/kanjirec (if public) — includes dataset generation, augmentation, training, evaluation, and ONNX export scripts.

v1 & v2 — Keras

KanjiDNN_v1.keras and KanjiDNN_v2.keras — the original Keras models.

Character classes

Kanji, Hiragana & Katakana (kana), Romaji (Latin letters), Digits (0-9), Common punctuation marks

Training data

Self-generated character images
SVG graphics from KanjiVG
ETL9G dataset from the ETL Character Database

Usage (Keras)

import os
os.environ["KERAS_BACKEND"] = "jax"  # or "torch", "tensorflow"
import keras

model = keras.saving.load_model("hf://gaiseras/kanjiDNN")

Try it live

Gradio demo on Hugging Face Spaces

Downloads last month: 95

gaiseras
/

kanjiDNN