basic_ocr.py

# coding: utf-8

# # Basic optical character using Tesseract with Python
# 
# 
# How to do a basic OCR for a PDF with incomplete or erroneous text information? For this a PDF of an Article is converted to an image and the tesseract package is used for the OCR. 
# 
# Tesseract is a popular free OCR-software https://sourceforge.net/projects/tesseract-ocr/. With a python wrapper it can be used to do OCR in python.

# Installation of additional python packages (assuming you have the basic anaconda packages installed):
# 
# #### basic OCR: install tesseract and opencv
# conda install opencv
# 
# conda install tesseract -c conda-forge
# 
# pip install pytesseract
# 
# #### convert pdf to img
# pip install pdf2image
# 
# #### read PDF
# conda install tika

# ## Example PDF file
# 
# As an example we use an article by A. Einstein which can be downloaded here: https://zenodo.org/record/1601163#.XsUtA7tR2EI as a PDF file. The file contains the scanned image of the article and text information. However, the text information contains errors. We will see if we can do better using the picture only and doing OCR with Tesseract.

# ## Read badly scanned PDF. 
# 
# #### First let's have a look at the PDF example

# In[1]:


from pdf2image import convert_from_path


# In[2]:


# read PDF as image
pages = convert_from_path('article.pdf', 500)


# In[3]:


# show first page of pdf as image
pages[0]


# ## For Comparison: read PDF using TIKA
# 
# This part is only to illustrate why we would be interested in optical character recognition for this example. This part can be skipped if you just want to know how to do the OCR. It requires the additional module tika for reading PDFs

# In[4]:


# import module for PDF reading
from tika import parser


# In[5]:


parsedPDF = parser.from_file('article.pdf')


# In[6]:


pdftext = parsedPDF['content']


# Now we look at the result below.

# In[7]:


print(pdftext)


# As you can see there are lots of mistakes!!! Please note that this is not an error in the TIKA-module. TIKA just reads whatever is stored as text in the PDF and does not do any OCR. For good PDFs it works perfectly

# # Here the OCR part  starts: 
# 
# First import all required packages

# In[8]:


# general tools
import os
import numpy as np


# In[9]:


# tools for OCR
import cv2
import pytesseract
from PIL import Image
from pytesseract import image_to_string


# ## Recognize characters from stored image
# 
# For this second part you need the PDF as image in JPG format. This can be done in various ways, e.g. using GIMP:https://www.gimp.org (Image handling and format conversion in python will be explained in a separate part of the toolbox.) 
# 
# In this first example we have stored the first page of the pdf as "Einstein1916_01.jpg"

# In[10]:


# define image name
img_name = "Einstein1916_01.jpg"
# define path where image is stored
src_path = './'


# In[11]:


# read image from file
img = cv2.imread(os.path.join(src_path,img_name))


# In[12]:


# now we use tesseract for character recognition
extxt = pytesseract.image_to_string(img)


# In[13]:


# Print result
print(extxt)


# This is slightly better, but there are still a lot of mistakes. Can we do better?

# ### Select language for character recognition
# 
# One reason for the bad quality is the fact that language-specific characters (German) are not recognized correctly. We can change this by loading the correct language data

# In[14]:


# tesseract character recognition with german language data
extxt = pytesseract.image_to_string(img, lang='deu')


# In[15]:


# Print result
print(extxt)


# Much better! This was the main Mistake! There are still some small errors, but the quality is good enough for text mining or stylometry.

# ## OCR for entire PDF + additional improvements 
# 
# Now we read all the pages (in this case just two) from the PDF. The OCR is slightly improved by adding some image conversion steps (this is not so important here, but may be useful in case of bad image quality). In the end we assemble the text from the PDF into one string, which could then be used e.g. to search for specific expressions.

# In[16]:


# first we read all the pdf pages (see above)
pages = convert_from_path('article.pdf', 500)


# In[17]:


# now we do image conversion and OCR for all pages
textpages = []
for n,page in enumerate(pages):
    open_cv_image = np.array(page)
    open_cv_image = open_cv_image[:, :, ::-1].copy() 
    # Convert to gray
    img = cv2.cvtColor(open_cv_image, cv2.COLOR_BGR2GRAY)
    # Apply dilation and erosion to remove some noise
    kernel = np.ones((1, 1), np.uint8)
    img = cv2.dilate(img, kernel, iterations=1)
    img = cv2.erode(img, kernel, iterations=1)
    # OCR, again using the german language set
    extxt = pytesseract.image_to_string(img,lang='deu')
    # Store Text
    textpages.append(extxt)


# In[18]:


# Assemble document text
text = textpages[0] + textpages[1]


# In[19]:


# Print document text
print(text)