0

I am trying to recognize handwritten text(capital letters and alphanumeric) and print it.

I have used the code mentioned in this link. handwriiten word recognition

I am also able to get different alphabets seperately in bounding boxes and it gives such results. enter image description here

I have two issues-

1 - I want to print out the text inside the bounding boxes

2- while writting continuously it considers the whole word as shown in the

output in github link but here it considers alphabets only.

here is the output for small letters continuous writting

enter image description here

can anyone provide me direction as to how I can do that with capital

letters and also how to print the text inside bounding boxes with good

accuracy so that it works good for capital letters alphanumeric.

Please anyone who could guide me regarding this and tell about how to do this.

Any help on handwritting recognition and text .

jatin singh
  • 123
  • 1
  • 1
  • 13

1 Answers1

1

Before finding the bounding box you will have to do some pre-processing.

I obtained a binary image using threshold from the image provided. After that I made a custom kernel such that it has more number of columns than rows to perform morphological operations. In this way, letters close to each other will be paired up.

Code:

custom_kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (30, 10))
threshed = cv2.morphologyEx(binary_image, cv2.MORPH_CLOSE, custom_kernel)
cv2.imshow('Connected letters', threshed)

enter image description here

After obtaining this you can find the bounding boxes.

Jeru Luke
  • 20,118
  • 13
  • 80
  • 87
  • Thanks for sharing such a valuable info and I will do the same. Can you just tell me how can I print the text inside the boxes so text gets extracted. – jatin singh May 24 '18 at 07:02
  • you can use the pytesseract library for text extraction! – Jeru Luke May 24 '18 at 07:06
  • 1
    for handwritten text recognition (HTR) you need a HTR model like CRNN (look at github) and train it on samples of your desired handwriting style and char-set. Maybe you can also retrain Tesseract. However, there is no out-of-the-box solution to recognize all possible handwriting styles that I am aware of. – Harry May 26 '18 at 14:47
  • @Harry is right. For handwritting text it's still not possible/very hard to have a solution. For computer/machine made text (like in a document/ID card/license plate/etc) you can use Tesseract (I suggest version 4). Also you can use MNIST for handwritting numbers along with a simple neural network which you can find in Google tutorials/GitHub. – lucians Jun 20 '18 at 13:30