Image-Processing-OCR-Walsh

This project contains a demonstration of using Walsh Functions to recognize characters in a clear image (without any noise, rotation or skew).

Requirements

• Python3
• Jupyter Notebook
• Matplotlib
• Numpy
• Opencv2
• Pillow
• Difflib

01 - Generate Walsh Functions

The file "jupyter/01-generate_walsh_functions.ipynb" contains the generation of 64 matrices obtained using Walsh Transformation and saved to a json file. Each is a 64x64 matrix.

02 - Build Database

The file "jupyter/02-build_database.ipynb" uses Pillow to generate a database that contains a set of characters, and a "Walsh Vector" for each character (consists of 64 values). Steps for each character:

• Generate blank Image, write the character on it
• Use vertical and horizontal segmentation to isolate the character
• Resize the image to 64x64
• Apply "Inner Product" between the image and each of the walsh matrices, resulting with 64 values that are considered as the character's "Walsh Vector"

The database is saved in a json file.

Note: some combinations of more than 1 character are considered as a whole character here, because they're very close to each other in Times font.

03 - Characters Segmentation

The file "jupyter/03-characters_segmentation.ipynb" uses Walsh matrices and the generated database to predict the values of the segmented characters in an image. Steps:

• Vertical Segmentation: gets the start and end pixel of each line.
• Horizontal Segmentaion for each line: gets the start and end pixel of each character in that line.
• Vertical Segmentation for each character: to remove spaces above & below the character
• Resize the character to 64x64
• Calculate the Walsh Vector for the character
• Search for the closest vector in the database using Manhattan Distance, and set the prediction to the character belonging to this vector