TrapRange: a Method to Extract Table Content in PDF Files

Source: http://www.dzone.com/articles/traprange-method-extract-table

Update

Utilize OpenAI API to extract information from PDF files
To extract information from PDF invoice
To run from the command line. Type java -jar traprange.latest.jar -h for help OR see examples in file test-command-line.sh

Introduction

Table data structure is one of the most important data structure in document, especially when exporting data from enterprise systems, data is usually in table format.

There are several data file formats that are often used to store tabular content such as CSV, text, and pdf. For the first two formats, it is quite straight forward just by opening file, loop through lines, and split cells with proper separator. The libraries to do this are quite a lot.

With PDF file, the story is completely different because it doesn't have a dedicated data definition for tabular content, something like table, tr, td tag in HTML. PDF is a complicated format with text data, font, styling, and also image, audio, and video, they can be mixed all together. Below is my proposed solution to data in high-density tabular content.

How to detect a table

After some investigation, I realized that:

Column: text content in cells of the same column lies on a rectangular space that does not overlap with other rectangular spaces of another column. For example, in the following image, red rectangle and blue rectangle are separated spaces
Row: words in the same horizontal alignment are in the same row. But this is just sufficient condition because a cell in a row may be a multi-line cell. For example, the fourth cell in the yellow rectangle has two lines, phrases “FK to this customer’s record in” and "Ledgers table" are not in the same horizontal alignment but they are still considered in the same row. In my solution, I simply assume that content in a cell only is single-line content. Different lines in a cell are considered to belong to different rows. Therefore the content in the yellow rectangle contains two rows: 1. {"Ledger_ID", "|", "Sales Ledger Account", "FK to this customer's record to"} 2. {NULL, NULL, NULL, "Ledgers table"}

PDFBox API

My library behind traprange is PDFBox which is the best PDF lib I know so far. To extract text from a pdf file, PDFBox API provides 4 classes:

PDDocument: contains information of entire pdf file. In order to load a pdf file, we use method PDDocument.load(stream: InputStream)
PDPage: represents each page in pdf document. We possibly archive a specific page content by passing the index of the page with this method: document.getDocumentCatalog().getAllPages().get(pageIdx: int)
TextPosition: represents an individual word or character in the document. We can fetch all TextPosition objects of a PDPage by overriding method processTextPosition(text: TextPosition) in class PDTextStripper. A TextPosition object has methods getX(), getY(), getWidth(), getHeight() that returns its bound in page and method getCharacter() to get its content.

In my work, I process text chunks directly by using TextPosition objects. For each text chunk in PDF file, it returns a text element with the following attributes:

x: horizontal distance from the left of the page
y: vertical distance from the top border of the page
maxX: equals x + width of the text chunk
maxY: equals y+ height of the text chunk

Trap ranges

The most important thing is identifying the bound of each row and column because if we know the bound of a row/column, we can retrieve all texts in that row/column from that we can easily extract all content inside the table and put it in a structured model. We name these bounds are trap-ranges. TrapRange has two attributes:

lowerBound: contains the lower endpoint of this range
upperBound: contains the upper endpoint of this range To calculate values of trap-ranges, we loop through all texts of the page and project range of each text onto the horizontal and vertical axis, get the result and join them together. After looping through all texts of the page, we will calculate trap-ranges and use them to identify cell data of the table.

Algorithm 1: calculating trap-ranges for each pdf page:

columnTrapRanges <-- []
rowTrapRanges <-- []
for each text in page
begin
     columnTrapRanges <-- join(columnTrapRanges, {text.x, text.x + text.width} )
     rowTrapRanges <-- join(rowTrapRanges, {text.y, text.y + text.height} )
end

After calculating trap-ranges for the table, we loop through all texts again and classify them into correct cells of the table.

Algorithm 2: classifying text chunks into correct cells:

table <-- new Table()
for each text in page
begin
     rowIdx <-- in rowTrapRanges, get index of the range that containts this text
     columnIdx <-- in columnTrapRanges, get index of the range that contains this text
     table.addText(text, rowIdx, columnIdx)
end

Design and implement

The above is class diagram of main classes:

TrapRangeBuilder: build() to calculate and return ranges
Table, TableRow and TableCell: for table data struture
PDFTableExtractor is the most important class. It contains methods to initialize and extract table data from PDF file. Builder pattern was used here. Following is some highlighted methods in this class:
setSource: set source of the pdf file. There're 3 overloads setSource(InputStream), setSource(File) and setSource(String)
addPage: to determine which pages will be processed. Default is all pages
exceptPage: to skip a page
exceptLine: to skip noisy data. All texts in these lines will be avoided.
extract: process and return result

Example

PDFTableExtractor extractor = new PDFTableExtractor();
List<Table> tables = extractor.setSource(“table.pdf”)
	.addPage(0)
	.addPage(1)
	.exceptLine(0) //the first line in each page
	.exceptLine(1) //the second line in each page
	.exceptLine(-1)//the last line in each page
	.extract();
String html = tables.get(0).toHtml();//table in html format
String csv = tables.get(0).toString();//table in csv format using semicolon as a delimiter

Following are some sample results (check out and run the test file TestExtractor.java):

Sample 1: Source: sample-1.pdf, result: sample-1.html
Sample 2: Source: sample-2.pdf, result: sample-2.html
Sample 3: Source: sample-3.pdf, result: sample-3.html
Sample 4: Source: sample-4.pdf, result: sample-4.html
Sample 5: Source: sample-5.pdf, result: sample-5.html

Evaluation

In experimentation, I used pdf files having high density of table content. The results show that my implementation detects tabular content better than other open-sources: pdftotext, pdftohtml, pdf2table. With documents having multi tables or too much noisy data, my method does not work well. If row has cells overlapped, columns of these cells will be merged.

Conclusion

TrapRange method works the best with PDF files having high density of table data. With documents have multi-table or too much noisy data, TrapRange is not a good choice. My method also can be implemented in other programming languages by replacing PDFBox by a corresponding pdf library or using command-line tool pdftohtml to extract text chunks and using these data as input data for algorithm 1, 2.

thoqbk/traprange