OCR_Persian_textDeceter

The purpose of this project

building and OCR model from the scratch to detect Persian text in pictures and turn it to a useful app in the future
creating the datasets and implementing the steps of Optical Character Detection system

Creating the Dataset

The first step in this project is to create the dataset. the dataset should compose of numbers of images with Persian text on them. To create these images:

first I used numpy library particulary the zeros function to create an blanck image :

img = np.zeros((300, 300, 3), np.uint8)

The np.zeros takes 2 argument to create blank images, the first argument is the shape of the array that we want to create, so we give the np.zeros function a tuple containing three numbers: width, height and the number of channels of image. And the second argument determines the type of the data of this array, for which we set unit8 (integer numbers). The function will return a 300*300 array filled with zeros represents a black blank image. The third number in the tuple represents the 3 RGB channels.

The next step was putting text on image:

Firstly I tried putting English text on the image by using OpenCV library. I used cv2.putText function to put the English text as it is shown below:

cv2.putText(img, 'Hejaz',(80, 150), font, 2, (255, 255, 255), 6, cv2.LINE_AA )

The function takes 8 arguments:

cv2.putText(image, the text, left-bottom coordinates of the text position , font, font-size, colour of the text, thikness of text, cv2.LINE_AA )

After putting the text on the image I use for loops to create numbers of images with English text on them.

But unfortunately, the opencv library does not contain any arabic or Persian fonts, and I was not able to install freetype-py, so I had to used pillow library to create the dataset.

Using PILLOW library to create image with Persian text

Creating blank images

to creat blank images we need firstly to import the Image module from pillow library, this module contains the new function that helped me to create blank images, this function takes 3 arguments :

img = Image.new('RGB', (300, 300), (255, 255, 255))

The arguments of Image.new are:

The mode of the image
The shape of image
The colour of the image

The colour is given as a tuple of 3 number each of the represent the red, the green and the blue channels of the images and each of them takes a number between 0, 255 the (255, 255, 255)represents the white colour.

Preparing the Persian text

After creating the blank images, the next step is to prepare the Persian text.

We first needed to install python-bidi and arabic_reshaper libraries, and then imported the arabic_reshaper library and the get_display function from bidi.algorithm.
We gave the Persian text as a string to reshape function of arabic_reshaper library, then the reshaped text was passed to get_display function.

Putting the Persian text on the image

Since the Pillow library has limited number of fonts and does not have any arabic or Persian fonts, we download an external Persian font and passed it to truetype function inside the ImageFont module of Pillow, to make us able to use it to put text on image and determine the size of the font as well.

To put the text we used Draw function of ImageDraw module and pass the blank image as an argument and assigned it to an object, then we used the text attribute of that object to put the text or word on the image:

''' draw = ImageDraw.Draw(img) draw.text((100, 150), bidi_word, font=font, fill='blue') img.save('img.png') '''

In the above mentioned text attribute we passed 4 parameters namely:

The left-top coordinate of the text position
The text we prepared
The font_style
The colour of the text

Then we saved the image by using save function on the img object, otherwise it will return a blank image.

One thing to be noticed here, is that in OpenCV library we had to pass the bottom-left coordinates of the text, but for the pillow library is the left-top coordinates.

Putting each of thses three stages in a function

The function of creating blank image
The function of preparing Persian/Arabic text
The function of putting text on image

Then put thses functions in 2 for loops to create the Dataset of images with Persian text.

For the text, I have used an excel file containing Persian words collected by REZA MOSHKSAR, converted it to csv file (I have used only 30000 words of the third column in the file), and then passed it as a list to the above mentioned loops to put each of them on seperated images.

To examine the code, I used matplotlib.pyplot library to visualize thses images in each steps.

Now the dataset is ready. It is a simple dataset contains 30000 images without any background noise and with just one Persian word on each of them.

The last step is to save these images to a directory. In order to perform this task, we first need to make a directory by using os module. After importing os module, we first create a directory named images_with_persian_text and join it to an exising path on Google drive ('/drive/My Drive/datasets') by using the code :

path = os.path.join(parent_path, directory)

Then pass the joined path to os.mkdir() function to make the directory. Now that the directory is ready, we use a for loop to save the images to this directory, the following code perform this:

for i, img in enumerate(image_text):
  img.save("{}/{}.png".format(path, i))

A sit is obvious in the code, we use the for loop to iterate over the list of created images, then apply the save function on each of img objects inside this list to save them to the directory. The string format function help us to make the saving path of these images by putting the path in the first placeholder {} and the name of the image in the second one. The mode of the image should be defined by putting an extention to the string, in this case we used png.

The next step is to train a RCNN with this dataset

We are going to create a RCNN network, split the data to train and test datasets, then we are going to train and evaluate the RCNN by these data.

hejazn86 / OCR_Persian_textDetecter