ParallelOCR
trying to speed up ocr using parallel computing in python
primarily optimising for scanned boooks
Steps
Create dataset
- list all fonts on google fonts with web api.
- select only Serif and Sans Serif Fonts.
- create template html file with all required letters.
- iterate over list of fonts and take screenshots of all letters separately using Multiple Selenium Processes.
- apply Preprocess on all letters and change resolution to 16x16 pixels.
Preprocess File
- get Scanned File.
- get page.
- apply transforms (process paralely if faster):
- grayscale (using otsu thresholding)
- contrast
- brightness
- sharpen edges
- compile File.
- or
- perform next step per page.
Seperate Lines
- find horizontal whitespace.
- divide image by horizontal whitespace.
- return cropped image of line.
OR - pass y coordinate of start and end of line.
crop image by line in next step. - assign index no to line.
- assign line to process.
- save processid with index no of line.
Recognize Text
- Tensorflow
- use CNN-RNN with CTC loss to train model on dataset .
- (similar to SimpleHTR but optimised for printed Text)
OR - use SVMs (because dataset would be relatively small)
- use multiprocessing to execute.
- blob detection for letters (group vertical groups)
- pass letters to trained model.
- Detect spaces.
- compile Text.
- return text.
Compile Page
- arrange returned text according to index and process id
- save as txt or pdf