Handwritten Chinese and Japanese OCR

This tutorial is also available as a Jupyter notebook that can be cloned directly from GitHub. See the installation guide for instructions to run this tutorial locally on Windows, Linux or macOS. To run without installing anything, click the launch binder button.

Binder Github

In this tutorial, we perform optical character recognition (OCR) for handwritten Chinese (simplified) and Japanese. An OCR tutorial using the Latin alphabet is available in notebook 208. This model is capable of processing only one line of symbols at a time.

The models used in this notebook are handwritten-japanese-recognition and handwritten-simplified-chinese. To decode model outputs as readable text kondate_nakayosi and scut_ept charlists are used. Both models are available on Open Model Zoo.


from collections import namedtuple
from itertools import groupby
from pathlib import Path

import cv2
import matplotlib.pyplot as plt
import numpy as np
from openvino.runtime import Core


Set up all constants and folders used in this notebook

# Directories where data will be placed
model_folder = "model"
data_folder = "data"
charlist_folder = f"{data_folder}/charlists"

# Precision used by model
precision = "FP16"

To group files, you have to define the collection. In this case, you can use namedtuple.

Language = namedtuple(
    typename="Language", field_names=["model_name", "charlist_name", "demo_image_name"]
chinese_files = Language(
japanese_files = Language(

Select Language

Depending on your choice you will need to change a line of code in the cell below.

If you want to do Japanese OCR, this line should be set to language = 'japanese' for Chinese set language = 'chinese'.

# Select language by using either language='chinese' or language='japanese'
language = "chinese"

languages = {"chinese": chinese_files, "japanese": japanese_files}

selected_language = languages.get(language)

Download Model

In addition to images and charlists, we need to download the model file. In the sections below there are cells for downloading either the Chinese or Japanese model.

If it is your first time running the notebook, the model will download. It may take a few minutes.

We use omz_downloader, which is a command-line tool from the openvino-dev package. omz_downloader automatically creates a directory structure and downloads the selected model.

path_to_model_weights = Path(f'{model_folder}/intel/{selected_language.model_name}/{precision}/{selected_language.model_name}.bin')
if not path_to_model_weights.is_file():
    download_command = f'omz_downloader --name {selected_language.model_name} --output_dir {model_folder} --precision {precision}'
    ! $download_command
omz_downloader --name handwritten-simplified-chinese-recognition-0001 --output_dir model --precision FP16
################|| Downloading handwritten-simplified-chinese-recognition-0001 ||################

========== Downloading model/intel/handwritten-simplified-chinese-recognition-0001/FP16/handwritten-simplified-chinese-recognition-0001.xml

========== Downloading model/intel/handwritten-simplified-chinese-recognition-0001/FP16/handwritten-simplified-chinese-recognition-0001.bin

Load Network and Execute

When all files are downloaded and language is selected, you need to read and compile the network to run inference. The path to the model is defined based on the selected language.

ie = Core()
path_to_model = path_to_model_weights.with_suffix(".xml")
model = ie.read_model(model=path_to_model)

Select Device Name

You may choose to run the network on multiple devices by default it will load the model on the CPU (you can choose manually CPU, GPU etc.) or let the engine choose the best available device (AUTO).

To list all available devices that you can use, uncomment and run the line print(ie.available_devices).

# To check available device names run the line below
# print(ie.available_devices)

compiled_model = ie.compile_model(model=model, device_name="CPU")

Fetch Information About Input and Output Layers

Now that the model is loaded, you need to fetch information about input and output layers. This is information about the input shape and the output.

recognition_output_layer = compiled_model.output(0)
recognition_input_layer = compiled_model.input(0)

Load an Image

The next step is to load an image.

The model expects a single-channel image as input, which is why we read the image in grayscale.

After loading the input image, the next step is getting information that you will use for calculating the scale ratio. This describes the ratio between required input layer height and the current image height. In the cell below, the image will be resized and padded to keep letters proportional and meet input shape.

# Read file name of demo file based on the selected model

file_name = selected_language.demo_image_name

# Text detection models expects an image in grayscale format
# IMPORTANT!!! This model allows to read only one line at time

# Read image
image = cv2.imread(filename=f"{data_folder}/{file_name}", flags=cv2.IMREAD_GRAYSCALE)

# Fetch shape
image_height, _ = image.shape

# B,C,H,W = batch size, number of channels, height, width
_, _, H, W = recognition_input_layer.shape

# Calculate scale ratio between input shape height and image height to resize image
scale_ratio = H / image_height

# Resize image to expected input sizes
resized_image = cv2.resize(
    image, None, fx=scale_ratio, fy=scale_ratio, interpolation=cv2.INTER_AREA

# Pad image to match input size, without changing aspect ratio
resized_image = np.pad(
    resized_image, ((0, 0), (0, W - resized_image.shape[1])), mode="edge"

# Reshape to network the input shape
input_image = resized_image[None, None, :, :]

Visualise Input Image

After preprocessing you can display the image.

plt.figure(figsize=(20, 1))
plt.imshow(resized_image, cmap="gray", vmin=0, vmax=255);

Prepare Charlist

The model is loaded and the image is ready. The only element left is the charlist which is downloaded, but before we use it, there is one more step. You must add a blank symbol at the beginning of the charlist. This is expected for both the Chinese and Japanese models.

# Get dictionary to encode output, based on model documentation
used_charlist = selected_language.charlist_name

# With both models, there should be blank symbol added at index 0 of each charlist
blank_char = "~"

with open(f"{charlist_folder}/{used_charlist}", "r", encoding="utf-8") as charlist:
    letters = blank_char + "".join(line.strip() for line in charlist)

Run Inference

Now run inference. compiled_model() takes a list with input(s) in the same order as model input(s). Then we can fetch the output from output tensors.

# Run inference on the model
predictions = compiled_model([input_image])[recognition_output_layer]

Process Output Data

The output of model format is W x B x L, where:

  • W - output sequence length

  • B - batch size

  • L - confidence distribution across the supported symbols in Kondate and Nakayosi.

To get a more human-readable format, select a symbol with the highest probability. When you hold a list of indexes that are predicted to have the highest probability, due to limitations in CTC Decoding, you will remove concurrent symbols and then remove the blanks.

The last step is getting the symbols from corresponding indexes in the charlist.

# Remove batch dimension
predictions = np.squeeze(predictions)

# Run argmax to pick the symbols with the highest probability
predictions_indexes = np.argmax(predictions, axis=1)
# Use groupby to remove concurrent letters, as required by CTC greedy decoding
output_text_indexes = list(groupby(predictions_indexes))

# Remove grouper objects
output_text_indexes, _ = np.transpose(output_text_indexes, (1, 0))

# Remove blank symbols
output_text_indexes = output_text_indexes[output_text_indexes != 0]

# Assign letters to indexes from output array
output_text = [letters[letter_index] for letter_index in output_text_indexes]