JeVois Tutorials  1.16
JeVois Smart Embedded Machine Vision Tutorials
Share this page:
Reading text with OpenCV and Tesseract

This tutorial develops a simplistic module that leverages optical character recognition (OCR) as supported by JeVois through OpenCV and the Tesseract library, which is available through the OpenCV Text module (including from Python).

Approach

  • We need to understand how to perform character recognition in OpenCV using Tesseract as a backend.
  • We will grab an image from the camera sensor, read text in it, and display the decoded text in a small area below the captured image. To achieve that, we will create a composite output display as in Creating composite output video frames

Creating the module

  • Select New Python Module... from the pull-down menu of JeVois Inventor (or press CTRL-N).
  • Fill in the details as shown below:

So, here we have decided to have a 60-pixel-tall area below the input image, where we will write the decoded text.

  • Allow JeVois to restart, select your new module from the Vision Module pull-down menu, and switch to the Code tab of the Inventor.

Writing the code

A bit of web search for OpenCV and Tesseract gets us to https://docs.opencv.org/3.4/d7/ddc/classcv_1_1text_1_1OCRTesseract.html which we will use to get our code going.

Basically, we need to create an OCR object (which we should do only once, hence we will do that in the constructor of our module), and then we will call run() on it with an image and a confidence level as inputs, getting decoded text (if any) as a result. We can finally send that text to serial and also write it into our output display.

First, let's get the OCR part done, just sending decoded text to serial port. Then we will worry about making the composite display later.

import libjevois as jevois
import cv2
import numpy as np
class TesseractOCR:
def __init__(self):
self.ocr = cv2.text.OCRTesseract_create()
def process(self, inframe, outframe):
img = inframe.getCvBGR()
txt = cv2.text_OCRTesseract.run(self.ocr, img, 50)
jevois.sendSerial('TXT {}'.format(txt))
outframe.sendCv(img)

Enable Module output to USB in the Console tab of the Inventor and... This is working!

Note
Note that the OpenCV OCR module is intended (as we understand it) to be run on small, pre-segmented boxes that each is likely to contain only one word and minimal surrounding garbage. This is usually accomplished by some scene text algorithm that is responsible for finding and extracting those boxes. Here, we bypass this step and just send the whole image to OCR. Because of this:
  • the algorithm is very slow as we process a whole 320x240 image
  • if multiple words are presented, the output shows them concatenated into one
  • if garbage is present, it may be decoded as text (as shown in the above log: often the garbage markings below the word Robotics are decoded into something, with varying success...

Adding the composite display output

We now just add a few lines of code to render the decoded text into a small message area below the video frame, and to stack both images up, as in Creating composite output video frames

import libjevois as jevois
import cv2
import numpy as np
class TesseractOCR:
def __init__(self):
self.ocr = cv2.text.OCRTesseract_create()
def process(self, inframe, outframe):
img = inframe.getCvBGR()
txt = cv2.text_OCRTesseract.run(self.ocr, img, 50)
jevois.sendSerial('TXT {}'.format(txt))
msgbox = np.zeros((60, img.shape[1], 3), dtype = np.uint8) + 80
cv2.putText(msgbox, txt, (3, 40), cv2.FONT_HERSHEY_SIMPLEX,
0.8, (255,255,255), 2, cv2.LINE_AA)
out = np.vstack((img, msgbox))
outframe.sendCv(out)

Here we go!

Going further

As mentioned above, better results will be obtained by implementing a pipeline that first extracts candidate boxes around each word, then sends each box to OCR (which can be parallelized over multiple boxes).

See for example https://github.com/opencv/opencv_contrib/blob/master/modules/text/samples/end_to_end_recognition.cpp (but it is written in C++, so some work will be needed to translate to python if desired).