node-tesseract-native

C++ module for node providing OCR with tesseract and leptonica

Prerequisites

Have linux or OSX (at least these are tested OSes)
Have node and node-gyp installed
Have leptonica (~1.68) libs and headers installed
Have tesseract (~3.02) libs and headers installed

Build

Checkout the repository and build it yourself using (you will also need this for the node10 branch)

node-gyp configure && node-gyp build

or use npm

npm install tesseract_native for nodejs > 0.12.0 <= 8.x.y

npm install tesseract_native@0.4.1 for nodejs >= 0.9.4 < 0.12.0

Supported Picture Formats

The module can handle every picture format leptonica can handle (see there), but as this module is likely to be used in an online service, pictures should be as small as possible. A 1.3 Megapixel picture converted to B/W using adaptive threshold filtering, saved as PNG will be 50KB on average. This is were you want to go.

Test your setup

You can test your setup using the provided test.js script on the command-line

$ node test.js HelloWorld.jpg

Example server

The code below shows a fully functional server where you can POST pictures to. The response will contain the recognized plain text or be empty if nothing was recognized or something went wrong.

var tesseract = require('tesseract_native');
var http = require('http');

var server = http.createServer(function(request, response)
{
    if(request.method === 'POST')
    {
        var totalSize = 0;
        var bufferList = new Array();
        var myOcr = new tesseract.OcrEio();
        
        request.on('data', function(data) {
            bufferList.push(data);
            totalSize += data.length;
            if (totalSize > 1e6) {
                console.log('Request body too large');
                request.connection.destroy();
            }
        });
        
        request.on('end', function() {
            var buffer = Buffer.concat(bufferList, totalSize);
            myOcr.ocr(buffer, function(err, result) {
                if(err) {
                    response.writeHead(500, {'Content-Type': 'text/plain'});
                    response.end("Error " + err);
                } else {
                    response.writeHead(200, {'Content-Type': 'text/plain'});
                    response.end(result);
                }
            });
        });
        
    } else {
        request.connection.destroy();
    }
}).listen(process.argv[2]);

Parameters

The OCR function also accepts a config object as second and the callback as third parameter like this:

myOcr.ocr(buffer, { lang:"eng", rect:[0,0,400,400] }, function(result) {
    // do something
});

The first supported parameter is tessdata, which is the path to you Tesseract data directory (/usr/local/share/tessdata/ by default). The second is lang which can be any three-character code for a language you have installed with Tesseract (eng by default). The third is rect, which is an array describing a rect of the form [X, Y, WIDTH, HEIGHT] limiting the image region for recognition. If you try the above rect with the provided test image it should land you in hell... Another parameter psm which takes an Integer from 0 to 10 configuring the page segmentation mode as in the table below. The default mode is 3.

Value	Meaning
0	Orientation and script detection (OSD) only.
1	Automatic page segmentation with OSD.
2	Automatic page segmentation, but no OSD, or OCR.
3	Fully automatic page segmentation, but no OSD. (Default)
4	Assume a single column of text of variable sizes.
5	Assume a single uniform block of vertically aligned text.
6	Assume a single uniform block of text.
7	Treat the image as a single text line.
8	Treat the image as a single word.
9	Treat the image as a single word in a circle.
10	Treat the image as a single character.

Why?

The question may arise. I've seen many tesseract wrappers for node and none of them I found did it quite right, some of them even did it wrong. The philosophy (and necessity) behind node is not to block, so everything that does work has to do it asynchronously and emit an event/execute a closure when it's done. If you don't do that, your node application will simply not perform well.

But even in this code you can see a very crude solution, performance-wise. The tesseract api is instantiated and initialized on every call to the ocr method. Why did't I do that when loading the module or when the constructor gets called? It has multiple reasons:

Simplicity: Initializing tesseract involves file system access, that means, must be performed asynchronously. The OCR work is done by adding the function to the asynchronous uv_queue_work, for simplicity I bundled all time-consuming tasks in the function that gets passed to the queue. So even though initializing wastes some cycles, it is still perfectly non-blocking.
Flexiblity: The language setting is passed on initialization, so by initializing on each request, the language used for detection can be set with each request.
Robustness: The tesseract context may not be thread-safe. There are hints in the tesseract code that suggest that. I will look further into it and I will be experimenting with a version that initializes the tesseract context at load time.

mdelete/node-tesseract-native