dedoc-utils

This library contains useful utilities for automatic document images processing:

1. Preprocessing
   • binarization
   • skew correction
2. Text detection
3. Line segmentation
4. Text recognition

Installation

The library requires Tesseract OCR to be installed. To install the library use the following command:

pip install dedoc-utils

It's supposed that you already have torch and torchvision installed. If not you can use the following command for installation:

pip install "dedoc-utils[torch]"

If you cloned the repository, you can install the dependencies via pip:

pip install .

To install torch packages use:

pip install ."[torch]"

Basic usage

Using preprocessors

from dedocutils.preprocessing import AdaptiveBinarizer, SkewCorrector
import cv2
import matplotlib.pyplot as plt

binarizer = AdaptiveBinarizer()
skew_corrector = SkewCorrector()

image = cv2.imread("examples/before_preprocessing.jpg")
binarized_image, _ = binarizer.preprocess(image)
preprocessed_image, _ = skew_corrector.preprocess(binarized_image)

fig = plt.figure(figsize=(10, 7))
rows, columns = 1, 3

fig.add_subplot(rows, columns, 1)
plt.imshow(image)
plt.axis('off')
plt.title("Before preprocessing")
  
fig.add_subplot(rows, columns, 2)
plt.imshow(binarized_image)
plt.axis('off')
plt.title("After binarization")

fig.add_subplot(rows, columns, 3)
plt.imshow(preprocessed_image)
plt.axis('off')
plt.title("After preprocessing")

Using text detectors

from dedocutils.text_detection import DoctrTextDetector

text_detector = DoctrTextDetector()
bboxes = text_detector.detect(preprocessed_image)

for bbox in bboxes[:5]:
    print(bbox)

BBox(x_top_left=2415, y_top_left=3730, width=202, height=97)
BBox(x_top_left=790, y_top_left=3613, width=383, height=105)
BBox(x_top_left=1690, y_top_left=3488, width=407, height=104)
BBox(x_top_left=2171, y_top_left=3488, width=377, height=92)
BBox(x_top_left=885, y_top_left=3505, width=27, height=50)

Using text recognizers

from dedocutils.text_recognition import TesseractTextRecognizer

text_recognizer = TesseractTextRecognizer()

for bbox in bboxes[:10]:
    word_image = preprocessed_image[bbox.y_top_left:bbox.y_bottom_right, bbox.x_top_left:bbox.x_bottom_right]
    text = text_recognizer.recognize(word_image, parameters=dict(language="eng"))
    print(text)

Fie-
afjefjores.
coluntur,
dicuntur
delubro
eodem
dii in
plures

Using line segmenters

In the previous example, the order of the recognized words isn't the same as the order of the words in the document. It happens because of undetermined work of the text detector. In this case, one may use line segmenter to sort bboxes from the text detector.

from dedocutils.line_segmentation import ClusteringLineSegmenter

line_segmenter = ClusteringLineSegmenter()
sorted_bboxes = line_segmenter.segment(bboxes)
for bbox in sorted_bboxes[1]:
    word_image = preprocessed_image[bbox.y_top_left:bbox.y_bottom_right, bbox.x_top_left:bbox.x_bottom_right]
    text = text_recognizer.recognize(word_image, parameters=dict(language="eng"))
    print(text)

gentes,
fimul.
obibant
munera
fumma
facra,