cosc428-structor
I had an open-ended Computer Vision assignment to complete, and an out-of-copyright book that I wanted to turn into an ebook. Conventional OCR engines like Tesseract weren't able to accurately recognise the page structure, which led to many transcription errors. If I could tell Tesseract to ignore certain regions (like images or repeated headers), then I could greatly reduce the number of errors in the resulting ebook. Thus: for my assignment, I wrote a program that takes an image and uses computer vision magick to determine the page's structure. So far, my program can detect and locate:
- lines of text,
- paragraphs,
- section titles,
- images and their associated captions,
- boilerplate like page numbers, and
- chapter titles.
Ain't it grand?
Dependencies
The project is written in Python 2.7.3 and uses the cv2 library for interacting with openCV. It also uses numpy for some of the mathematical operations. On windows, the best way to get these dependencies is to install the Python(x,y) suite (https://code.google.com/p/pythonxy/), which combines python with a customisable set of scientific computing libraries.
Program Structure
The program's root is main.py, but this simply iterates through images in a folder and constructs a Page instance from each image. Thus, the real work happens in page.py.
page.py contains a few utility methods and the Page class. The constructor calls the appropriate methods in order to determine the logical structure of the page. This structure is stored in three objects: self.margin, self.content, and self.boilerplate (which contains such non-content text objects as the page number and header).
The getBuildingBlocks method is responsible for finding words, grouping words into textual lines, discarding marginal noise, and fitting a Margin instance around the remaining lines. Most of these tasks are preformed by calling other functions.
The self.content object is found by passing the set of lines to the Content() constructor. This uses a state machine to group lines into figures, paragraphs, section titles, etc. The Content class, along with a class for each content type, is found in content.py.
The other files can generally be ignored when trying to understand the program; they are largely just convenience classes which represent page elements (such as points, geometric lines, words, text lines, and boxes), as well as supporting tools such as the Stopwatch.
How to Run the Code
Run main.py using the python interpreter. This will process each page in ./images, and for each page a series of 'snapshot' images will be displayed in order to illustrate the algorithm. To show only the final result for each image, set showSteps in main.py to False.
3 Feb 17, 2022
15 Nov 9, 2022
1.5k Dec 28, 2022
1.2k Dec 29, 2022
186 Dec 29, 2022
71 Dec 6, 2022
99 Dec 12, 2022
59 Sep 10, 2022
422 Jan 3, 2023
2.5k Jan 03, 2023
67 Nov 30, 2022
17 Nov 22, 2022
140 Nov 04, 2022
170 Dec 21, 2022
239 Dec 13, 2022
160 Dec 21, 2022
170 Dec 27, 2022
182 Dec 29, 2022
459 Dec 26, 2022
2.2k Jan 04, 2023
29 Dec 26, 2022
20 Jan 06, 2023
684 Jan 06, 2023
1.3k Jan 04, 2023
0 Nov 01, 2021
48 Jan 02, 2023
17 Dec 08, 2022
27 Jun 20, 2022