WORK IN PROGRESS
A Clojure application that extracts matrixes from PDF's files and saves those as csv files.
A separate application extracts the word positions from a PDF.
Copyright © 2014 FIXME
Distributed under the Eclipse Public License either version 1.0 or (at your option) any later version.
How it works
The input file contains the positions from the PDF. The format of a single word position:
- page number
- line number
- font name
- font size
- space width
- x1 (x1,y1) - (x2,y2) are the bounding box of the word
The font information is currently not used, but will serve to estimate the minimum width / height.
The list of words is sorted by page / line / x1. For each page/line, the spans are calculated (the white space between the words) The small spans are removed (minimum width as given as input parameter)
The spans are transformed to stripes. Spans are defined on a single line, in constrast stripes will cover multiple lines. If a span of a line N overlaps a span of the line N+1, the common part of the overlapping spans will form a stripe. A stripe will be stretched as long there is an overlapping span in the next line. The stripes are vertical separators without crossing a word.
The process of building stripes is repeated for each line. Line 1 will have stripes that start on that line and descend as far as possible. Line 2 will have stripes that start on line 2 and descend as far as possible,...
The small stripes are removed (not wide or height enought, defined as input parameter)
A column is defined as the space between 2 stripes. The shortest stripe defines the height of the column. Like the stripes, each line will have a number of columns that start on that line.
For each column, a simple statistic 'word-fillage' is calculated, which reflects the number of words in that column relative to the number of lines covered by that column. (The idea is to prefer full columns)
Again, a number of small columns are removed (almost-empty columns).
The goal is to give a score to each line. The score indicates how 'good' the columns are. The current implementation uses a simple score: the score is the size of the matrix (columns x height).
Smarter strategies could try to merge several columns, or use the height of the columns per line etc...
The content of the best grid is stored as a csv file.