OpenCV bindings for Node.js. OpenCV is the defacto computer vision library - by interfacing with it natively in node, we get powerful real time vision in js.
People are using node-opencv to fly control quadrocoptors, detect faces from webcam images and annotate video streams. If you're using it for something cool, I'd love to hear about it!
You'll need OpenCV 2.3.1 or newer installed before installing node-opencv.
Install OpenCV using brew
brew install pkg-config
brew install opencv@2
brew link --force opencv@2
Download and install OpenCV (Be sure to use a 2.4 version) @ http://opencv.org/releases.html For these instructions we will assume OpenCV is put at C:\OpenCV, but you can adjust accordingly.
If you haven't already, create a system variable called OPENCV_DIR and set it to C:\OpenCV\build\x64\vc12
Make sure the "x64" part matches the version of NodeJS you are using.
Also add the following to your system PATH ;%OPENCV_DIR%\bin
Install Visual Studio 2013. Make sure to get the C++ components. You can use a different edition, just make sure OpenCV supports it, and you set the "vcxx" part of the variables above to match.
Download peterbraden/node-opencv fork git clone https://github.com/peterbraden/node-opencv
run npm install
$ npm install opencv
Run the examples from the parent directory.
cv.readImage("./examples/files/mona.png", function(err, im){
im.detectObject(cv.FACE_CASCADE, {}, function(err, faces){
for (var i=0;i<faces.length; i++){
var x = faces[i]
im.ellipse(x.x + x.width/2, x.y + x.height/2, x.width/2, x.height/2);
}
im.save('./out.jpg');
});
})
The matrix is the most useful base data structure in OpenCV. Things like images are just matrices of pixels.
new Matrix(rows, cols)
Or if you're thinking of a Matrix as an image:
new Matrix(height, width)
Or you can use opencv to read in image files. Supported formats are in the OpenCV docs, but jpgs etc are supported.
cv.readImage(filename, function(err, mat){
...
})
cv.readImage(buffer, function(err, mat){
...
})
If you need to pipe data into an image, you can use an ImageDataStream:
var s = new cv.ImageDataStream()
s.on('load', function(matrix){
...
})
fs.createReadStream('./examples/files/mona.png').pipe(s);
If however, you have a series of images, and you wish to stream them into a stream of Matrices, you can use an ImageStream. Thus:
var s = new cv.ImageStream()
s.on('data', function(matrix){
...
})
ardrone.createPngStream().pipe(s);
Note: Each 'data' event into the ImageStream should be a complete image buffer.
var mat = new cv.Matrix.Eye(4,4); // Create identity matrix
mat.get(0,0) // 1
mat.row(0) // [1,0,0,0]
mat.col(3) // [0,0,0,1]
mat.save('./pic.jpg')
or:
var buff = mat.toBuffer()
im.convertGrayscale()
im.canny(5, 300)
im.houghLinesP()
im.ellipse(x, y)
im.line([x1,y1], [x2, y2])
There is a shortcut method for Viola-Jones Haar Cascade object detection. This can be used for face detection etc.
mat.detectObject(haar_cascade_xml, opts, function(err, matches){})
For convenience in face detection, cv.FACE_CASCADE is a cascade that can be used for frontal face detection.
Also:
mat.goodFeaturesToTrack
mat.findCountours
mat.drawContour
mat.drawAllContours
findContours
returns a Contours
collection object, not a native array. This object provides
functions for accessing, computing with, and altering the contours contained in it.
See relevant source code and examples
var contours = im.findContours();
// Count of contours in the Contours object
contours.size();
// Count of corners(verticies) of contour `index`
contours.cornerCount(index);
// Access vertex data of contours
for(var c = 0; c < contours.size(); ++c) {
console.log("Contour " + c);
for(var i = 0; i < contours.cornerCount(c); ++i) {
var point = contours.point(c, i);
console.log("(" + point.x + "," + point.y + ")");
}
}
// Computations of contour `index`
contours.area(index);
contours.arcLength(index, isClosed);
contours.boundingRect(index);
contours.minAreaRect(index);
contours.isConvex(index);
contours.fitEllipse(index);
// Destructively alter contour `index`
contours.approxPolyDP(index, epsilon, isClosed);
contours.convexHull(index, clockwise);
It requires to train
then predict
. For acceptable result, the face should be cropped, grayscaled and aligned, I ignore this part so that we may focus on the api usage.
Please ensure your OpenCV 3.2+ is configured with contrib. MacPorts user may port install opencv +contrib
const fs = require('fs');
const path = require('path');
const cv = require('opencv');
function forEachFileInDir(dir, cb) {
let f = fs.readdirSync(dir);
f.forEach(function (fpath, index, array) {
if (fpath != '.DS_Store')
cb(path.join(dir, fpath));
});
}
let dataDir = "./_training";
function trainIt (fr) {
// if model existe, load it
if ( fs.existsSync('./trained.xml') ) {
fr.loadSync('./trained.xml');
return;
}
// else train a model
let samples = [];
forEachFileInDir(dataDir, (f)=>{
cv.readImage(f, function (err, im) {
// Assume all training photo are named as id_xxx.jpg
let labelNumber = parseInt(path.basename(f).substring(3));
samples.push([labelNumber, im]);
})
})
if ( samples.length > 3 ) {
// There are async and sync version of training method:
// .train(info, cb)
// cb : standard Nan::Callback
// info : [[intLabel,matrixImage],...])
// .trainSync(info)
fr.trainSync(samples);
fr.saveSync('./trained.xml');
}else {
console.log('Not enough images uploaded yet', cvImages)
}
}
function predictIt(fr, f){
cv.readImage(f, function (err, im) {
let result = fr.predictSync(im);
console.log(`recognize result:(${f}) id=${result.id} conf=${100.0-result.confidence}`);
});
}
//using defaults: .createLBPHFaceRecognizer(radius=1, neighbors=8, grid_x=8, grid_y=8, threshold=80)
const fr = new cv.FaceRecognizer();
trainIt(fr);
forEachFileInDir('./_bench', (f) => predictIt(fr, f));
Using tape. Run with command:
npm test
.
I (@peterbraden) don't spend much time maintaining this library, it runs primarily on contributor support. I'm happy to accept most PR's if the tests run green, all new functionality is tested, and there are no objections in the PR.
Because I haven't got much time for maintenance, I'd prefer to keep an absolute minimum of dependencies.
The library is distributed under the MIT License - if for some reason that doesn't work for you please get in touch.