How to improve Hindi text extraction?

Question

I am trying to extract Hindi text from a PDF. I tried all the methods to exract from the PDF, but none of them worked. There are explanations why it doesn't work, but no answers as such. So, I decided to convert the PDF to an image, and then use pytesseract to extract texts. I have downloaded the Hindi trained data, however that also gives highly inaccurate text.

That's the actual Hindi text from the PDF (download link):

That's my code so far:

import fitz

filepath = "D:\\BADI KA BANS-Ward No-002.pdf"

doc = fitz.open(filepath)
page = doc.loadPage(3)  # number of page
pix = page.getPixmap()
output = "outfile.png"
pix.writePNG(output)
from PIL import Image
import pytesseract

# Include tesseract executable in your path
pytesseract.pytesseract.tesseract_cmd = r"C:\Program Files\Tesseract-OCR\tesseract.exe"

# Create an image object of PIL library
image = Image.open('outfile.png')

# pass image into pytesseract module
# pytesseract is trained in many languages
image_to_text = pytesseract.image_to_string(image, lang='hin')

# Print the text
print(image_to_text)

That's some output sample:

कार बिता देवी व ०... नाम बाइुनान िक०क नाक तो
पति का नाव: रवजी लात. “50९... पिला का सामशामाव.... “पति का नाम: बादुलल
कान सब: 43 लसमनंध्या: 93९. मकान ंब्या: 3९
आप: 29 _ लिंग सी. | आइ 57 लिंग पुरुष आप: 62 लिंग सी
एजगल्णब्णस्य (बन्द जगाख्मिणण्य
नमः बायगी बसों ०४... नि बयावर्णो ०५०... निफर सनक नी
चिता का नामजबूजल वर्ष.“ ००० | पिला का नामब्राइलाल वर्षो... 0 2... | पिता कामामशुल चब्द .... “20०
|सकानसंब्या: 43९ बसवकंब्या: 43९. कान संब्या: 44
जाए: 27 लिंग सो कई: 27 नि खी मा लिंग पुरुष

There is an answer to this question I want to scrape a Hindi(Indian Langage) pdf file with python, which seems to tell how to do this, but provides no explanation whatsoever.

Is there any way to do this?

Answer 1

I will give some ideas how to process your image, but I will limit that to page 3 of the given document, i.e. the page shown in the question.

For converting the PDF page to some image, I used pdf2image.

For the OCR, I use pytesseract, but instead of lang='hin', I use lang='Devanagari', cf. the Tesseract GitHub. In general, make sure to work through Improving the quality of the output from the Tesseract documentation, especially the page segmentation method.

Here's a (lengthy) description of the whole procedure:

1. Inverse binarize the image for contour finding: white texts, shapes, etc. on black background.
2. Find all contours, and filter out the two very large contours, i.e. these are the two tables.
3. Extract texts outside of the two tables:
   1. Mask out tables in the binarized image.
   2. Do morphological closing to connect remaining lines of text.
   3. Find contours, and bounding rectangles of these lines of text.
   4. Run pytesseract to extract the texts.
4. Extract texts inside of the two tables:
   1. Extract the cells, better: their bounding rectangles, from the current table.
   2. For the first table:
      1. Run pytesseract to extract the texts as-is.
   3. For the second table:
      1. Floodfill the rectangle around the number to prevent faulty OCR output.
      2. Mask the left (Hindi) and right (English) part.
      3. Run pytesseract using lang='Devaganari' on the left, and using lang='eng' on the right part to improve OCR quality for both.

That'd be the whole code:

import cv2
import numpy as np
import pdf2image
import pytesseract

# Extract page 3 from PDF in proper quality
page_3 = np.array(pdf2image.convert_from_path('BADI KA BANS-Ward No-002.pdf',
                                              first_page=3, last_page=3,
                                              dpi=300, grayscale=True)[0])

# Inverse binarize for contour finding
thr = cv2.threshold(page_3, 128, 255, cv2.THRESH_BINARY_INV)[1]

# Find contours w.r.t. the OpenCV version
cnts = cv2.findContours(thr, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)
cnts = cnts[0] if len(cnts) == 2 else cnts[1]

# STEP 1: Extract texts outside of the two tables

# Mask out the two tables
cnts_tables = [cnt for cnt in cnts if cv2.contourArea(cnt) > 10000]
no_tables = cv2.drawContours(thr.copy(), cnts_tables, -1, 0, cv2.FILLED)

# Find bounding rectangles of texts outside of the two tables
no_tables = cv2.morphologyEx(no_tables, cv2.MORPH_CLOSE, np.full((21, 51), 255))
cnts = cv2.findContours(no_tables, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)
cnts = cnts[0] if len(cnts) == 2 else cnts[1]
rects = sorted([cv2.boundingRect(cnt) for cnt in cnts], key=lambda r: r[1])

# Extract texts from each bounding rectangle
print('\nExtract texts outside of the two tables\n')
for (x, y, w, h) in rects:
    text = pytesseract.image_to_string(page_3[y:y+h, x:x+w],
                                       config='--psm 6', lang='Devanagari')
    text = text.replace('\n', '').replace('\f', '')
    print('x: {}, y: {}, text: {}'.format(x, y, text))

# STEP 2: Extract texts from inside of the two tables

rects = sorted([cv2.boundingRect(cnt) for cnt in cnts_tables],
               key=lambda r: r[1])

# Iterate each table
for i_r, (x, y, w, h) in enumerate(rects, start=1):

    # Find bounding rectangles of cells inside of the current table
    cnts = cv2.findContours(page_3[y+2:y+h-2, x+2:x+w-2],
                            cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)
    cnts = cnts[0] if len(cnts) == 2 else cnts[1]
    inner_rects = sorted([cv2.boundingRect(cnt) for cnt in cnts],
                         key=lambda r: (r[1], r[0]))

    # Extract texts from each cell of the current table
    print('\nExtract texts inside table {}\n'.format(i_r))
    for (xx, yy, ww, hh) in inner_rects:

        # Set current coordinates w.r.t. full image
        xx += x
        yy += y

        # Get current cell
        cell = page_3[yy+2:yy+hh-2, xx+2:xx+ww-2]

        # For table 1, simply extract texts as-is
        if i_r == 1:
            text = pytesseract.image_to_string(cell, config='--psm 6',
                                               lang='Devanagari')
            text = text.replace('\n', '').replace('\f', '')
            print('x: {}, y: {}, text: {}'.format(xx, yy, text))

        # For table 2, extract single elements
        if i_r == 2:

            # Floodfill rectangles around numbers
            ys, xs = np.min(np.argwhere(cell == 0), axis=0)
            temp = cv2.floodFill(cell.copy(), None, (xs, ys), 255)[1]
            mask = cv2.floodFill(thr[yy+2:yy+hh-2, xx+2:xx+ww-2].copy(),
                                 None, (xs, ys), 0)[1]

            # Extract left (Hindi) and right (English) parts
            mask = cv2.morphologyEx(mask, cv2.MORPH_CLOSE,
                                    np.full((2 * hh, 5), 255))
            cnts = cv2.findContours(mask,
                                    cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)
            cnts = cnts[0] if len(cnts) == 2 else cnts[1]
            boxes = sorted([cv2.boundingRect(cnt) for cnt in cnts],
                           key=lambda b: b[0])

            # Extract texts from each part of the current cell
            for i_b, (x_b, y_b, w_b, h_b) in enumerate(boxes, start=1):

                # For the left (Hindi) part, extract Hindi texts
                if i_b == 1:

                    text = pytesseract.image_to_string(
                        temp[y_b:y_b+h_b, x_b:x_b+w_b],
                        config='--psm 6',
                        lang='Devanagari')
                    text = text.replace('\f', '')

                # For the left (English) part, extract English texts
                if i_b == 2:

                    text = pytesseract.image_to_string(
                        temp[y_b:y_b+h_b, x_b:x_b+w_b],
                        config='--psm 6',
                        lang='eng')
                    text = text.replace('\f', '')

                print('x: {}, y: {}, text:\n{}'.format(xx, yy, text))

And, here are the first few lines of the output:

Extract texts outside of the two tables

x: 972, y: 93, text: राज्य निर्वाचन आयोग, राजस्थान
x: 971, y: 181, text: पंचायत चुनाव निर्वाचक नामावली, 2021
x: 166, y: 610, text: मिश्र का बाढ़ ,श्रीराम की नॉगल
x: 151, y: 3417, text: आयु 1 जनवरी 2021 के अनुसार
x: 778, y: 3419, text: पृष्ठ संख्या : 3 / 10

Extract texts inside table 1

x: 146, y: 240, text: जिलापरिषद का नाम : जयपुर
x: 1223, y: 240, text: जि° प° सदस्य निर्वाचन क्षेत्र : 21
x: 146, y: 327, text: पंचायत समिति का नाम : सांगानेर
x: 1223, y: 327, text: पं° स° सदस्य निर्वाचन क्षेत्र : 6
x: 146, y: 415, text: ग्रामपंचायत : बडी का बांस
x: 1223, y: 415, text: वार्ड क्रमांक : 2
x: 146, y: 502, text: विधानसभा क्षेत्र की संख्या एवं नाम:- 56-बगरु

Extract texts inside table 2

x: 142, y: 665, text:
1 RBP2469583
नाम: आरती चावला
पिता का नामःलाला राम चावला
मकान संख्याः १९
आयुः 21 लिंगः स्त्री

x: 142, y: 665, text:
Photo is
Available

x: 867, y: 665, text:
2 MRQ3101367
नामः सूरज देवी
पिता का नामःरामावतार
मकान संख्याः डी /18
आयुः 44 लिंगः स्त्री

x: 867, y: 665, text:
Photo is
Available

I checked a few texts using manual character-by-character comparison, and thought it looked quite good, but unable to understand Hindi or reading Devanagari script, I can't comment on the overall quality of the OCR. Please let me know!

Annoyingly, the number 9 from the corresponding "card" is falsely extracted as 2. I assume, that happens due to the different font compared to the rest of the text, and in combination with lang='Devanagari'. Couldn't find a solution for that – without extracting the rectangle separately from the "card".

----------------------------------------
System information
----------------------------------------
Platform:      Windows-10-10.0.19041-SP0
Python:        3.9.1
PyCharm:       2021.1.1
NumPy:         1.19.5
OpenCV:        4.5.2
pdf2image      1.14.0
pytesseract:   5.0.0-alpha.20201127
----------------------------------------

Answer 2

if you want to scrape 100% correct text from pdf, you should use the correct font family and encode at the parsing time from image to text.

Answer 3

If you want to get texts from these 'cards' I've managed to do it for page 3 via module tabula-py this way:

import tabula

pdf_file = "BADI KA BANS-Ward No-002.pdf"
page = 3

x = 30      # left edge of the table
y = 160     # top edge of the table
w = 173     # width of a card
h = 73      # height of a card
photo = 61  # width of a photo

rows = 8    # number of rows of the table
cols = 3    # number of columns of the table

counter = 1

def get_area(row, col):
    ''' return area of the card in given position in the table '''
    top    = y + h * row
    left   = x + w * col
    bottom = top + h
    right  = left + w - photo
    return (top, left, bottom, right)

for row in range(rows):
    for col in range(cols):
        file_name = "card_" + str(counter).zfill(3) + ".txt"
        tabula.convert_into(pdf_file, file_name,
        pages=page,
        output_format = "csv",
        java_options = "-Dfile.encoding=UTF8",
        lattice = False,
        area = get_area(row, col))
        counter += 1

Input:

Output

24 txt files:

card_001.txt
card_002.txt
card_003.txt
card_004.txt
.
.
.
card_023.txt
card_024.txt

card_001.txt:

1 RBP2469583
नरम: आरतल चररलर
नपतर कर नरम:लरलर ररम चररल
मकरन सखजर: १९
आज:  21 ललग: सल

card_002.txt

2 MRQ3101367
नरम: सरज दरल
नपतर कर नरम:ररमररतरर
मकरन सखजर: रल /18
आज:  44 ललग: सल

card_024.txt

24 RBP0230979
नरम: सनमतकरर
पनत कर नरम: हररलसह
मकरन सखजर: 13
आज:  41 ललग: सल

As far as I can see all the 'cards' have the same dimensions. The solution could be applied for all pages if they look alike. Unfortunately the pages have differences. So initial variables must be changed for every page. I see no way to make the changes automatically. Except the numbers of the cards can be taken from the card instead of the simply counter.

https://pypi.org/project/tabula-py/

https://aegis4048.github.io/parse-pdf-files-while-retaining-structure-with-tabula-py