Some test text!

menu
search
chevron_right C++ samples

C++ use OCR to make searchable PDFs and extract text

Sample C++ code shows how to use the PDFTron OCR module on scanned documents in multiple languages. The OCR module can make searchable PDFs and extract scanned text for further indexing.

To run this sample, get started with a free trial of PDFTron SDK.

//---------------------------------------------------------------------------------------
// Copyright (c) 2001-2019 by PDFTron Systems Inc. All Rights Reserved.
// Consult legal.txt regarding legal and license information.
//---------------------------------------------------------------------------------------
#include <PDF/PDFNet.h>
#include <PDF/PDFDoc.h>
#include <PDF/OCRModule.h>
#include <PDF/OCROptions.h>
#include <SDF/SDFDoc.h>
#include <SDF/Obj.h>
#include <iostream>

using namespace std;
using namespace pdftron;
using namespace PDF;
using namespace SDF;

//---------------------------------------------------------------------------------------
// The following sample illustrates how to use OCR module
//---------------------------------------------------------------------------------------
int main(int argc, char *argv[])
{
	try 
	{
		// The first step in every application using PDFNet is to initialize the 
		// library and set the path to common PDF resources. The library is usually 
		// initialized only once, but calling Initialize() multiple times is also fine.
		PDFNet::Initialize();
		// The location of the OCR Module
		PDFNet::AddResourceSearchPath("../../../Lib/");
		if(!OCRModule::IsModuleAvailable())
		{
			cout << endl;
			cout << "Unable to run OCRTest: PDFTron SDK OCR module not available." << endl;
			cout << "---------------------------------------------------------------" << endl;
			cout << "The OCR module is an optional add-on, available for download" << endl;
			cout << "at http://www.pdftron.com/. If you have already downloaded this" << endl;
			cout << "module, ensure that the SDK is able to find the required files" << endl;
			cout << "using the PDFNet::AddResourceSearchPath() function." << endl << endl;
			return 0;
		}

		// Relative path to the folder containing test files.
		string input_path =  "../../TestFiles/OCR/";
		string output_path = "../../TestFiles/Output/";

		//--------------------------------------------------------------------------------
		// Example 1) Process image without specifying options, default language - English - is used
		try 
		{

			// A) Setup empty destination doc

			PDFDoc doc;

			// B) Run OCR on the .png with options

			OCRModule::ImageToPDF(doc.GetSDFDoc(), input_path + "psychomachia_excerpt.png", NULL);

			// C) check the result

			doc.Save(output_path + "psychomachia_excerpt.pdf", 0, 0);
			
			cout << "Example 1: psychomachia_excerpt.png" << endl;

		}
		catch(Common::Exception& e)	
		{
			cout << e << endl;
		}
		catch(...) 
		{
			cout << "Unknown Exception" << endl;
		}

		//--------------------------------------------------------------------------------
		// Example 2) Process document using multiple languages
		try
		{
			// A) Setup empty destination doc

			PDFDoc doc;

			// B) Setup options with multiple target languages, English will always be considered as secondary language

			OCROptions opts;
			RectCollection rects;
			opts.AddLang("rus");
			opts.AddLang("deu");

			// C) Run OCR on the .jpg with options

			OCRModule::ImageToPDF(doc.GetSDFDoc(), input_path + "multi_lang.jpg", &opts);

			// D) check the result

			doc.Save(output_path + "multi_lang.pdf", 0, 0);

			cout << "Example 2: multi_lang.jpg" << endl;

		}
		catch (Common::Exception& e)
		{
			cout << e << endl;
		}
		catch (...)
		{
			cout << "Unknown Exception" << endl;
		}

		//--------------------------------------------------------------------------------
		// Example 3) Process a .pdf specifying a language - German - and ignore zone comprising a sidebar image 
		try
		{
			// A) Open the .pdf document

			SDFDoc doc((input_path + "german_kids_song.pdf").c_str());

			// B) Setup options with a single language and an ignore zone

			OCROptions opts;
			opts.AddLang("deu");

			RectCollection ignore_zones;
			ignore_zones.AddRect(1768, 680, 2056, 3044);
			opts.AddIgnoreZonesForPage(ignore_zones, 1);

			// C) Run OCR on the .pdf with options

			OCRModule::ProcessPDF(doc, &opts);

			// D) check the result

			PDFDoc doc_out(doc);
			doc_out.Save(output_path + "german_kids_song.pdf", 0, 0);

			cout << "Example 3: german_kids_song.pdf" << endl;
		}
		catch (Common::Exception& e)
		{
			cout << e << endl;
		}
		catch (...)
		{
			cout << "Unknown Exception" << endl;
		}

		//--------------------------------------------------------------------------------
		// Example 4) Process multipage tiff with text/ignore zones specified for each page, optionally provide English as the target language
		try
		{
			// A) Setup empty destination doc

			PDFDoc doc;

			// B) Setup options with a single language plus text/ignore zones

			OCROptions opts;
			opts.AddLang("eng");

			RectCollection ignore_zones;
			// ignore signature box in the first 2 pages
			ignore_zones.AddRect(1492, 56, 2236, 432);
			opts.AddIgnoreZonesForPage(ignore_zones, 1);

			ignore_zones.Clear();
			ignore_zones.AddRect(1492, 56, 2236, 432);
			opts.AddIgnoreZonesForPage(ignore_zones, 2);

			// can use a combination of ignore and text boxes to focus on the page area of interest,
			// as ignore boxes are applied first, we remove the arrows before selecting part of the diagram
			ignore_zones.Clear();
			ignore_zones.AddRect(992, 1276, 1368, 1372);
			opts.AddIgnoreZonesForPage(ignore_zones, 3);


			RectCollection text_zones;
			// we only have text zones selected in page 3

			// select horizontal BUFFER ZONE sign
			text_zones.AddRect(900, 2384, 1236, 2480);
			// select right vertical BUFFER ZONE sign
			text_zones.AddRect(1960, 1976, 2016, 2296);
			// select Lot No.
			text_zones.AddRect(696, 1028, 1196, 1128);

			// select part of the plan inside the BUFFER ZONE
			text_zones.AddRect(428, 1484, 1784, 2344);
			text_zones.AddRect(948, 1288, 1672, 1476);
			opts.AddTextZonesForPage(text_zones, 3);

			// C) Run OCR on the .tif with options

			OCRModule::ImageToPDF(doc, input_path + "bc_environment_protection.tif", &opts);

			// D) check the result

			doc.Save(output_path + "bc_environment_protection.pdf", 0, 0);

			cout << "Example 4: bc_environment_protection.tif" << endl;

		}
		catch (Common::Exception& e)
		{
			cout << e << endl;
		}
		catch (...)
		{
			cout << "Unknown Exception" << endl;
		}

		//--------------------------------------------------------------------------------
		// Example 5) Alternative workflow for extracting OCR result JSON, postprocessing (e.g., removing words not in the dictionary or filtering special
		// out special characters), and finally applying modified OCR JSON to the source PDF document 
		try
		{

			// A) Open the .pdf document

			SDFDoc doc((input_path + "zero_value_test_no_text.pdf").c_str());

			// B) Run OCR on the .pdf with default English language

			UString json = OCRModule::GetOCRJsonFromPDF(doc, NULL);

			// C) Post-processing step (whatever it might be)

			cout << "Have OCR result JSON, re-applying to PDF " << endl;

			// D) Apply potentially modified OCR JSON to the PDF

			OCRModule::ApplyOCRJsonToPDF(doc, json);

			// E) Check the result

			PDFDoc doc_out(doc);
			doc_out.Save(output_path + "zero_value_test_no_text.pdf", 0, 0);

			cout << "Example 5: extracting and applying OCR JSON from zero_value_test_no_text.pdf" << endl;

		}
		catch (Common::Exception& e)
		{
			cout << e << endl;
		}
		catch (...)
		{
			cout << "Unknown Exception" << endl;
		}

		//--------------------------------------------------------------------------------
		// Example 6) The postprocessing workflow has also an option of extracting OCR results in XML format, similar to the one used by TextExtractor
		try
		{

			// A) Setup empty destination doc

			PDFDoc doc;

			// B) Run OCR on the .tif with default English language, extracting OCR results in XML format. Note that
			// in the process we convert the source image into PDF. We reuse this PDF document later to add hidden text layer to it.

			UString xml = OCRModule::GetOCRXmlFromImage(doc, input_path + "physics.tif", NULL);

			// C) Post-processing step (whatever it might be)

			cout << "Have OCR result XML, re-applying to PDF" << endl;

			// D) Apply potentially modified OCR XML to the PDF

			OCRModule::ApplyOCRXmlToPDF(doc, xml);

			// E) Check the result

			PDFDoc doc_out(doc);
			doc_out.Save(output_path + "physics.pdf", 0, 0);

			cout << "Example 6: extracting and applying OCR XML from physics.tif" << endl;

		}
		catch (Common::Exception& e)
		{
			cout << e << endl;
		}
		catch (...)
		{
			cout << "Unknown Exception" << endl;
		}

		cout << "Done." << endl;

		PDFNet::Terminate();
	}
	catch(Common::Exception& e)	
	{
		cout << e << endl;
	}
	catch (...) {
		cout << "Unknown Exception" << endl;
	}

	return 0;	
}
close

Free Trial

Get unlimited trial usage of PDFTron SDK to bring accurate, reliable, and fast document processing capabilities to any application or workflow.

Select a platform to get started with your free trial.

Unlimited usage. No email address required.

PDFTron Receives USD$71 Million Growth Investment Led By Silversmith Capital Partners

Learn more
close