Build your own Optical Character Recognition (OCR) System using Google’s Tesseract and OpenCV

Aniruddha Bhandari 28 Nov, 2020 • 9 min read

Overview

  • Optical Character Recognition (OCR) is a widely used system in the computer vision space
  • Learn how to build your own OCR for a variety of tasks
  • We will leverage the OpenCV library and Tesseract for building the OCR system

 

Introduction

Do you remember the days when you had to fill in the dots of the right answer during an exam? Or how about the aptitude test you gave before your first job? I can vividly recall the olympiads and multiple-choice tests where universities and organizations used an Optical Character Recognition (OCR) system to grade the answer sheets in droves.

Honestly, OCR has applications in a broad range of industries and functions. So, everything from scanning documents – bank statements, receipts, handwritten documents, coupons, etc.,  to reading street signs in autonomous vehicles – this all falls under the OCR umbrella.

OCR systems used to be quite expensive and cumbersome to build a couple of decades ago. But advances in the computer vision and deep learning field mean we can build our own OCR system right now!

OCR_Tesseract

But building an OCR system isn’t a straightforward task. For starters, it is filled with problems like different fonts in images, poor contrast, multiple objects in an image, etc.

So, in this article, we will explore some very famous and effective approaches for the OCR task and how you can implement one yourself.

If you are new to object detection and computer vision, I suggest going through the following resources:

 

Table of Contents

  1. What is Optical Character Recognition (OCR)?
  2. Popular OCR Applications in the Real World
  3. Text Recognition with Tesseract OCR
  4. The Different Ways for Text Detection

 

What is Optical Character Recognition (OCR)?

Let’s first understand what OCR is, in case you haven’t come across this concept before.

OCR, or Optical Character Recognition, is a process of recognizing text inside images and converting it into an electronic form. These images could be of handwritten text, printed text like documents, receipts, name cards, etc., or even a natural scene photograph.

OCR has two parts to it. The first part is text detection where the textual part within the image is determined. This localization of text within the image is important for the second part of OCR, text recognition, where the text is extracted from the image. Using these techniques together is how you can extract text from any image.

OCR But nothing is perfect and OCR is no exception. However, with the advent of deep learning, it has become possible to get better and more generalized solutions to this problem.

Before we dive into how to build your own OCR, let’s take a look at some of the popular applications of OCR.

 

Popular OCR Applications in the Real World

OCR has widespread applications across industries (primarily with the aim of reducing manual human effort). It has been incorporated in our everyday life to an extent that we hardly ever notice it! But they surely strive to bring a better user experience.

ocr_applicationsOCR is used for handwriting recognition tasks to extract information. A lot of work is going on in this field and we have made some really significant advancements. Microsoft has come up with an awesome mathematical application that takes as input a handwritten mathematical equation and generates the solution along with a step-by-step explanation of the working.

OCR is increasingly being used for digitization by various industries to cut down manual workload. This makes it very easy and efficient to extract and store information from business documents, receipts, invoices, passports, etc. Also, when you upload your documents for KYC (Know Your Customer), OCR is used to extract information from these documents and store them for future reference.

OCR is also used for book scanning where it turns raw images into a digital text format. Many large scale projects like the Gutenberg project, Million Book Project, and Google Books use OCR to scan and digitize books and store the works as an archive.

The banking industry is also increasingly using OCR to archive client-related paperwork, like onboarding material, to easily create a client repository. This significantly reduces the onboarding time and thereby improves the user experience. Also, banks use OCR to extract information like account number, amount, cheque number from cheques for faster processing.

OCR_banking

The applications of OCR are incomplete without mentioning their use in self-driving cars. Autonomous cars rely extensively on OCR to read signposts and traffic signs. An effective understanding of these signs makes autonomous cars safe for pedestrians and other vehicles that ply on the roads.

There are definitely many more applications of OCR like vehicle number plate recognition, converting scanned documents into editable word documents, and many more. I would love to hear your experience of using OCR – let me know in the comments section below.

The digitization using OCR obviously has widespread advantages like easy storage and manipulation of the text, not to mention the unfathomable amount of analytics that you can apply to this data! OCR is definitely one of the most important fields of Computer Vision.

Now, let’s look at one of the most famous and widely used text recognition techniques – Tesseract.

 

Text Recognition with Tesseract OCR

Tesseract is an open-source OCR engine originally developed as proprietary software by HP (Hewlett-Packard) but was later made open source in 2005. Google has since then adopted the project and sponsored its development.

Tesseract

As of today, Tesseract can detect over 100 languages and can process even right-to-left text such as Arabic or Hebrew! No wonder it is used by Google for text detection on mobile devices, in videos, and in Gmail’s image spam detection algorithm.

From version 4 onwards, Google has given a significant boost to this OCR engine. Tesseract 4.0 has added a new OCR engine that uses a neural network system based on LSTM (Long Short-term Memory), one of the most effective solutions for sequence prediction problems. Although its previous OCR engine using pattern matching is still available as legacy code.

Once you have downloaded Tesseract onto your system, you easily run it from the command line using the following command:

tesseract <test_image> <output_file_name> -l <language(s)> --oem <mode> --psm <mode> 

You can change the Tesseract configuration for results best suited for your image:

  1. Langue (-l) – You can detect a single language or multiple languages with Tesseract
  2. OCR engine mode (–oem) – As you already know, Tesseract 4 has both LSTM and Legacy OCR engines. However, there are 4 modes of valid operation modes based on their combination

OCR engine

3. Page Segmentation (–psm) – Can be adjusted according to the text in the image for better results

OCR page segmentation

 

Pyteseract

However, instead of the command-line method, you could also use Pytesseract – a Python wrapper for Tesseract. Using this you can easily implement your own text recognizer using Tesseract OCR by writing a simple Python script.

You can download Pytesseract using the pip install pytesseract command.

The main function in Pytesseract is image_to_text() which takes the image and the command line options as its arguments:

Pytesseract

 

What are the Challenges with Tesseract?

It’s no secret that Tesseract is not perfect. It performs poorly when the image has a lot of noise or when the font of the language is one on which Tesseract OCR is not trained. Other conditions like brightness or skewness of text will also affect the performance of Tesseract. Nevertheless, it is a good starting point for text recognition with low efforts and high outputs.

 

The Different Ways for Text Detection

Tesseract assumes that the input text image is fairly clean. Unfortunately, many input images will contain a plethora of objects and not just a clean preprocessed text. Therefore, it becomes imperative to have a good text detection system that can detect text which can then be easily extracted.

There are a fair few ways for text detection:

  • Traditional way of using OpenCV
  • Contemporary way of using Deep Learning models, and
  • Building your very own custom model

 

Text Detection using OpenCV

Text detection using OpenCV is the classic way of doing things. You can apply various manipulations like image resizing, blurring, thresholding, morphological operations, etc. to clean the image.

Image processing

Here we have Grayscale, blurred and thresholded images, in that order.

Once you have done that, you can use OpenCV contours detection to detect contours to extract chunks of data:

Finally, you can apply text recognition on the contours that you got to predict the text:

OCR OpenCV

The results in the image above were achieved with minimum preprocessing and contour detection followed by text recognition using Pytesseract. Obviously, the contours did not detect the text every time.

But, still, doing text detection with OpenCV is a tedious task requiring a lot of playing around with the parameters. Also, it does not do well in terms of generalization. A better way of doing this is by using the EAST text detection model.

 

Contemporary Deep Learning Model – EAST

EAST, or Efficient and Accurate Scene Text Detector, is a deep learning model for detecting text from natural scene images. It is pretty fast and accurate as it is able to detect 720p images at 13.2fps with an F-score of 0.7820.

The model consists of a Fully Convolutional Network and a Non-maximum suppression stage to predict a word or text lines. The model, however, does not include some intermediary steps like candidate proposal, text region formation, and word partition that were involved in other previous models, which allows for an optimized model.

You can have a look at the image below provided by the authors in their paper comparing the EAST model with other previous models:

Text detection models

EAST has a U-shape network. The first part of the network consists of convolutional layers trained on the ImageNet dataset. The next part is the feature merging branch which concatenates the current feature map with the unpooled feature map from the previous stage.

This is followed by convolutional layers to reduce computation and produce output feature maps. Finally, using a convolutional layer, the output is a score map showing the presence of text and a geometry map which is either a rotated box or a quadrangle that covers the text. This can be visually understood from the image of the architecture that was included in the research paper:

EAST architechture

I highly suggest you go through the paper yourself to get a good understanding of the EAST model.

OpenCV has included the EAST text detector model in version 3.4 onwards. This makes it super convenient to implement your own text detector. The resulting localized text boxes can be passed through Tesseract OCR to extract the text and you will have a complete end-to-end model for OCR.

EAST application

 

Custom Model using TensorFlow Object API for Text Detection

The final method to build your text detector is using a custom-built text detector model using the TensorFlow Object API. It is an open-source framework used to build deep learning models for object detection tasks. To understand it in detail, I suggest going through this detailed article first.

To build your custom text detector, you would obviously require a dataset of quite a few images, at least more than 100. Then you need to annotate these images so that the model can know where the target object is and learn everything about it. Finally, you can choose from one of the pre-trained models, depending on the trade-off between performance and speed, from TensorFlow’s detection model zoo. You can refer to this comprehensive blog to build your custom model.

Now. training can require some computation, but if you don’t really have enough of it, don’t worry! You can use Google Colaboratory for all your requirements! This article will teach you how to use it effectively.

Finally, if you want to go a step ahead and build a YOLO state-of-the-art text detector model, this article will be a stepping stone to understanding all the nitty-gritty of it and you will be off to a great start!

 

End Notes

In this article, we covered the problems in OCR and the various approaches that can be used to solve the task. We also discussed the various shortcomings in the approaches and why OCR is not as easy as it seems!

Have you worked with any OCR application before? What kind of OCR use cases do you plan on building after this? Let me know your ideas and feedback below.

Frequently Asked Questions

Lorem ipsum dolor sit amet, consectetur adipiscing elit,

Responses From Readers

Clear

Andrew Cameron Morris
Andrew Cameron Morris 16 May, 2020

Back in 2018 I tried to use Tesseract, together with OpenCV, to read text from large tables. However, when the tables were very dense the table box lines interfered a lot with the recognition, resulting in unacceptably low recognition accuracy. After quite some effort I was able to detect each table, and then each cell within the table. I could then pass each cell separately to Tesseract for text recognition. This resulted in much greater recognition accuracy. I was also able to reassemble the recognised text data into a pandas DataFrame, which is vital if the table data needs to be automatically associated with the table row and/or column headers. While the python tool I created to do this is proprietary, I may be able to help if anyone is interested in tips on how to go about doing this.

Ahmed ALLALI
Ahmed ALLALI 12 Jun, 2020

Hello Andrew, really nice to have your feedback, Thanks ! I'm working on text extraction from scanned administrative forms, it contains handwritten and printed text and a bit complex tables. The problem with handwritten text extraction is that it relies heavily on the form of the text, so it's crucial to estimate accurately the performance of this extraction. It would be really helpful if I have a deeper information about your experience. Thank you in advance !

Shyam
Shyam 28 Jun, 2020

Hi Andrew, good to note you are keen to help. I need to develop scanned bill document processing module which will be able to extract free text and table data irrespective of the different format in which the data is presented. My customer don't want to retrain for every new format or template. Is there a way to achieve this ? please share your view. Thanks, Shyam.

Kamal
Kamal 31 Dec, 2020

To anyone who reads this message: Dear all of you, I BEG YOU: INVENT AN EFFECTIVE OPTICAL RECOGNITION SOLUTION FOR THE ARABIC LANGUAGE TO RECOGNIZE 100% OF THE VOCALIZED TEXT (WITH THE CORRECT SIGNS ABOVE AND BELOW EACH LETTER) ... I TESTED so many softwares that I can't count them anymore : The best at the moment (to my limited knowledge): - Easy screen OCR (a software that works with a Google API), - Sakhr (a software from the Alalamiah group) which costs between 2000€ and 5000€ (depending on the version), - Abbyy Finereader 15... FRANKLY ALL OF THEM PROMISE YOU A RECOGNITION OF MORE THAN 90% BUT IT IS FAR FROM BEING THE CASE (COUNT BETWEEN 50 AND 60% EVEN AFTER HAVING IMPROVED THE IMAGE AS MUCH AS POSSIBLE FOR OCHRE)... FOR A SIMPLE 200-WORD PAGE: I SPEND MORE THAN AN HOUR CORRECTING EVERYTHING... IT'S AWFUL!!! THANK YOU FOR YOUR GENEROSITY: PLEASE PUT YOUR INTELLIGENCE AT THE SERVICE OF HUMANITY! KINDLY BEST REGARDS,

vInnovate Technologies
vInnovate Technologies 09 Jan, 2022

EAST (Efficient and Accurate Scene Text Detector) is a deep learning model for detecting text from images of natural scenes. vDigiDocr is cloud-based document-driven OCR software that can run AI to automate low-value business tasks and digitize business flows.