#FCV_Project3

---

Assumptions for handwritten character extractions:

• The color used is in contrast to the written surface so that a thresholding can pe safely applied to extract written text. For example, black or blue ink on white paper
• White paper or light colored paper is used
• The written lines are relatively parallel to each other

Algorithm used for line extraction:

The following steps are taken in order to extract lines of text:

• Transform the image in a binary image
• Find the rectangle with the minimum area that contains all the "white" points (the written text)
• Apply a rotation based on the found rectangle above
• Calculate the average point value for each line (the percentage of pixels corresponding to text) in the warped image. An example of this histogram can be seen below:
• Find all the lines for which there is a change (raising or falling) in percentage

Algorithm used for character extraction:

The following steps are taken in order to extract handwritten characters:

• Transform the warped image in a gray image
• Apply Canny edge detection algorithm
• Find all the contours based on the edges
• Find the corresponding bounding box (which is big enough to remove the noise) and draw it

Different approaches that were tried:

• Using EAST Neural Network to detect words:
  • No way to extract individual characters.
  • Relatively poor performance for handwritten words, especially "wavy" lines of text

Usage of software:

From command line, run the code like so: python main.py <img_path_1> <img_path_2> <...>
The SW will write the processed image in the /data/output folder adding processed_ in front of the image name