Jurajzovinec / homework-nemocnica-one

Optical character recognition homework

---

APPLICATION URL https://ocr-recognition-homework.herokuapp.com/

Note this is prototype and has some issues: Issues:

• long boot time
• lack of operational memory may lead to 503 status code or fail of server
• AWS limit

---

This ocr application uses Otsu algorithm in order to preprocess image file. Application uses AWS s3 storage as temporarly file storage system and allows user to get thresholded image with DOWNLOAD IMAGE button.

Web app workflow

• User initiate app with UPLOAD CAR SIGNS IMAGES
• Uploaded images are sent to backend
• Backend REST API stores images are stored on AWS S3 storage system
• process main.py is spawned from node
• result of this process is sent to fronend, where is visualized as table row

Python part

Spawned and top layer script is main.py which returns result of its process by printing in json object. For the purpose of this application have been coded three modules:

• aws_functions.py -> which consinsts of functions regarding communication with AWS S3 file storage cloud
• image_manipulation_functions.py.py -> which consists of functions regarding image processing
• validate_car_sign.py -> 'business logic' with no external dependecies, evaluates scanned text from image with real life czech evidential car sign patterns

For thresholding has been used OTSU algorithm which has been imported from pytreshold package

otsu.py (pytreshold package)

def otsu_threshold(image=None, hist=None):
    """ Runs the Otsu threshold algorithm.

    Reference:
    Otsu, Nobuyuki. "A threshold selection method from gray-level
    histograms." IEEE transactions on systems, man, and cybernetics
    9.1 (1979): 62-66.

    @param image: The input image
    @type image: ndarray
    @param hist: The input image histogram
    @type hist: ndarray

    @return: The Otsu threshold
    @rtype int
    """
    if image is None and hist is None:
        raise ValueError('You must pass as a parameter either'
                         'the input image or its histogram')

    # Calculating histogram
    if not hist:
        hist = np.histogram(image, bins=range(256))[0].astype(np.float)

    cdf_backg = np.cumsum(np.arange(len(hist)) * hist)
    w_backg = np.cumsum(hist)  # The number of background pixels
    w_backg[w_backg == 0] = 1  # To avoid divisions by zero
    m_backg = cdf_backg / w_backg  # The means

    cdf_foreg = cdf_backg[-1] - cdf_backg
    w_foreg = w_backg[-1] - w_backg  # The number of foreground pixels
    w_foreg[w_foreg == 0] = 1  # To avoid divisions by zero
    m_foreg = cdf_foreg / w_foreg  # The means

    var_between_classes = w_backg * w_foreg * (m_backg - m_foreg) ** 2

    return np.argmax(var_between_classes)

Setup application (windows)

prerequisites:

• Node +v14.16.1
• python +3.9.0
• npm +6.14.12
• tesseract

1. Init python environment with Command Prompt:

python -m venv python_packages && python_packages\Scripts\activate && pip install -r requirements.txt

2. Select local environment with Powershell:

. python_packages\Scripts\activate

3. Set local path to tesseract.exe in module:

def read_text_from_image(image, config):

    load_dotenv()

    # Development
    pytesseract.pytesseract.tesseract_cmd = os.getenv('PATH_TO_TESSERACT_EXE')

    # Production (Heroku)
    # pytesseract.pytesseract.tesseract_cmd = '/app/.apt/usr/bin/tesseract'

    read_text = pytesseract.image_to_string(image, config=config)

    sanitized_text = remove_non_alphanumeric_chars(read_text)

    return sanitized_text

4. Install dependecies for Node (backend) with Powershell/Command Prompt:

npm i

5. Install dependecies for React (frontend) Command Prompt:

cd client && npm i

6. Start backend with Powershell/Command Prompt:

npm run start

7. Start frontend with Command Prompt:

cd client && npm run start

8. Create .env file with valid keys:

AWS_ACCESS_KEY=yourAWSAccessKey
AWS_SECRET_KEY=youtAWSSecretKey
AWS_BUCKET_NAME=car-sing-parking-recognition-bucket
PATH_TO_TESSERACT_EXE='C:\\Program Files\\Tesseract-OCR\\tesseract.exe'