Tesseract-OCR, Python, Computer Vision

Question

I want to train tesseract for my own new fonts but I am not finding any way to do it. I have trouble in creating box files from images. I am new to programming languages and some one told me about LabelImg but it is not useful for tesseract ocr.

Please help me suggesting a tool to label text from image which are new to tesseract ocr.

Answer 1

You can create your own script to label images. Here is some sample code that will allow you to do so, you can customize it if you want

import sys
import os

import cv2


def isImage(filepath) -> bool:
    '''
    checks if file is an image
    '''

    lowercasePath = filepath.lower()

    # you can add more formats here
    cases = [
        lowercasePath.endswith('jpg'),
        lowercasePath.endswith('png'),
        lowercasePath.endswith('jpeg'),
    ]

    return any(cases)



def getPaths(imgdir, condition=lambda x: True):
    '''
    given path to image folder will return you a list of full paths
    to files which this folder contain

    :param condition: is a function that will filter only those files
    that satisfy condition
    '''

    files = map(lambda x: os.path.join(imgdir, x).strip(),
        os.listdir(imgdir))

    filtered = filter(condition, files)

    return list(filtered)



def labelingProcess(imgdir):
    print("Welcome to the labeling tool")
    print("if you want to stop labeling just close the program or press ctrl+C")

    WIDTH = 640
    HEIGHT = 480

    WINDOWNAME = "frame"
    window = cv2.namedWindow(WINDOWNAME, cv2.WINDOW_NORMAL)
    cv2.resizeWindow(WINDOWNAME, WIDTH, HEIGHT)
    cv2.moveWindow(WINDOWNAME, 10, 10)


    pathsToImages = getPaths(imgdir, isImage)

    if not len(pathsToImages):
        print("couldn't find any images")
        return

    for pathtoimage in pathsToImages:
        imageName = os.path.basename(pathtoimage)

        # label img has the same name as image only ends with .txt
        labelName = ''.join(imageName.split('.')[:-1]) + '.gt.txt'
        labelPath = os.path.join(imgdir, labelName)

        # skip labeled images
        if os.path.exists(labelPath):
            continue

        # read image
        image = cv2.imread(pathtoimage)
        if image is None:
            print("couldn't open the image")
            continue

        h, w = image.shape[:2]

        # resize to fixed size (only for visualization)
        hnew = HEIGHT
        wnew = int(w * hnew / h)

        image = cv2.resize(image, (wnew, hnew))

        cv2.imshow(WINDOWNAME, image)
        cv2.waitKey(1)

        print("enter what is written on the image or \
              press enter to skip or")
        label = input()

        if not len(label):
            continue


        with open(labelPath, 'w') as labelfile:
            labelfile.write(label)

    cv2.destroyAllWindows()


if __name__ == '__main__':
    imgdir = sys.argv[1]
    labelingProcess(imgdir)

for this particular script requirement is opencv

usage:

python3 labelingtool.py <path to your folder with images>

It will read images from from your folder and create corresponding .gt.txt file with annotation. During labelling process you can type annotation in the terminal.

Further to train your own model you can use for example this repo https://github.com/thongvm/ocrd-train

It needs dataset to be in the format image and corresponding annotation

image1.tif
image1.gt.txt 

image2.tif
image2.gt.txt 

...

To convert your images to .tif you can use mogrify for example

this code will convert all jpg files to tif files

mogrify -format tif *.jpg