apsisocr

apsis-ocr

Apsis-OCR is a Mixed language ocr system for Printed Documents developed at Apsis Solutions limited

The full system is build with 3 components:

• Text detection : DBNet
• Text recognition:
  • Bangla Text : ApsisNet
    • ApsisNet is a model developed at Apsis Solutions Limited.
    • It is used by bbOCR as the recognition model
    • ApsisNet is found to be the best among other available recognition models (such as tesseract and easyOCR) in the linked paper
  • English Text : SVTR-LCNet
• Text classification : DenseNet121

Installation

As module/pypi package

cpu installation

pip install apsisocr
pip install onnxruntime
pip install fastdeploy-python -f https://www.paddlepaddle.org.cn/whl/fastdeploy.html

gpu installation

It is recommended to use conda environment . Specially for GPU.

• installing cudatoolkit and cudnn:

conda install cudatoolkit
conda install cudnn

• installing packages

pip install apsisocr
pip install onnxruntime-gpu
python -m pip install -U fastdeploy-gpu-python -f https://www.paddlepaddle.org.cn/whl/fastdeploy.html

• exporting environment variables

mkdir -p $CONDA_PREFIX/etc/conda/activate.d
echo 'export LD_LIBRARY_PATH=$CUDNN_PATH/lib:$CONDA_PREFIX/lib/:$LD_LIBRARY_PATH' >> $CONDA_PREFIX/etc/conda/activate.d/env_vars.sh

Building from source : Linux/Ubuntu

It is recommended to use conda environment .

• clone the repository :

git clone https://github.com/mnansary/apsisOCR.git
cd apsisOCR

• create a conda environment:

conda create -n apsisocr python=3.9

• activate conda environment:

conda activate apsisocr

• cpu installation :

bash install.sh cpu

• gpu installation :

bash install.sh gpu

Useage

Apsisnet : Bangla Recognizer

• useage

from apsisocr import ApsisNet
bnocr=ApsisNet()
bnocr.infer(crops)

• docstring for ApsisNet.infer

"""
Perform inference on image crops.

Args:
    crops (list[np.ndarray]): List of image crops.
    batch_size (int): Batch size for inference (default: 32).
    normalize_unicode (bool): Flag to normalize unicode (default: True).

Returns:
    list[str]: List of inferred texts.
"""

SVTR-LCNet : English Recognizer

from apsisocr import SVTRLCNet
enocr=SVTRLCNet()
enocr.infer(crops)

• docstring for SVTRLCNet.infer

"""
Perform inference on image crops.

Args:
    crops (list[np.ndarray]): List of image crops.
    batch_size (int): Batch size for inference.

Returns:
    list[str]: List of recognized texts.
"""

DenseNet121BnEnClassifier : Language classifier

from apsisocr import DenseNet121BnEnClassifier
lang=DenseNet121BnEnClassifier()
lang.infer(crops)

• docstring for DenseNet121BnEnClassifier.infer

"""
Perform inference on image crops.

Args:
    crops (list[np.ndarray]): List of image crops.
    batch_size (int): Batch size for inference (default: 32).

Returns:
    list[str]: List of inferred languages.
"""

PaddleDBNet : Text Detector

• check paddleOCR official website for better understanding of the model

# initialization
from apsisocr import PaddleDBNet
detector=PaddleDBNet()
# getting word boxes
word_boxes=detector.get_word_boxes(img)
# getting line boxes
line_boxes=detector.get_line_boxes(img)
# getting crop with either of the results
crops=detector.get_crops(img,word_boxes)

ApsisOCR : Overall System

from apsisocr import ApsisOCR
ocr=ApsisOCR()
results=ocr(img_path)

• docstring for ApsisOCR.__call__

"""
Perform OCR on an image.

Args:
    img_path (str): Path to the image file.

Returns:
    dict: OCR results containing recognized text and associated information. The dictionary has the following structre
            {
            "text" : multiline text with newline separators
            "result" : list a dictionaries that contains the following structre:
                        {
                        "line_no" : the line number of the word
                        "word_no" : the word number in the line 
                        "poly"    : the four point polygonal bounding box of the word in the image
                        "text"    : the recognized text 
                        "lang"    : the classified language code
                        }
            }
"""  

check useage/useage.ipynb for examples

Deployment

• cd deployment: change directory to deployment folder
• change the configs as required in config.py

# This port will be used by the api_ocr.py 
OCR_API_PORT=3032
# This api address is diplayed after deploying api_ocr.py and this is used in app.py  
OCR_API="http://172.20.4.53:3032/ocr"

• running the api and app:

python api_ocr.py # deploys at port 3032 by defautl
streamlit run app.py --server.port 3033 # deploys streamlit built frontend at 3033 port

• The api_ocr.py lets the api to be used without any UI (a postman screenshot is attached below)

• The app.py runs a streamlit UI

TESTED GPU INFERENCE SERVER CONFIG

OS          : Ubuntu 20.04.6 LTS      
Memory      : 62.4 GiB 
Processor   : Intel® Xeon(R) Silver 4214R CPU @ 2.40GHz × 24    
Graphics    : NVIDIA RTX A6000/PCIe/SSE2
Gnome       : 3.36.8

License

Contents of this repository are restricted to non-commercial research purposes only under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0).

Change Log

0.0.1 (24/09/2023)

• added useage and examples

0.0.2 (24/09/2023)

• added liscencing

0.0.3 (05/11/2023)

• added base application without line and word number

0.0.4 (01/05/2024)

• apsisbnocr

0.0.5 (22/11/2024)

• added Null handling to apsisbnocr

0.0.6 (22/11/2024)

• added Null handling to apsisbnocr - output typo update

0.0.7 (22/11/2024)

• BNbaseOCR -> location and text only