Crop Face Images

[safi099]

For better recognition I cropped the face of the person. https://github.com/safi099/Face_Recognition/blob/master/app/src/main/java/com/developer/iamsafi/face_recognition/FaceOverLay.java

Firstly I convert the Mat mGray to Bitmap and then perform operation to crop the face and there for the images Array I again convert into Mat format of mGray using the following code. mGray = new Mat(scale.getHeight(), scale.getWidth(), CvType.CV_8U); Bitmap bmp32 = scale.copy(Bitmap.Config.ARGB_8888, true); Utils.bitmapToMat(bmp32, mGray) Images is properly cropped but when I train the error comes the Image must be 8 bit matrix from the native class train class. Please Help me to sort out the problem

[safi099]

Also Tell me about the eigen faces that how many training images are sufficent and what value of face threshold and distance threshold it will give better results.

[Lauszus]

My guess would be that the created bitmap has more than one channel, so when the you convert it back to a Mat object, then it is not a grayscale image.

Try this code instead:

if (scale != null) {
    Mat origMat = new Mat();
    Bitmap bmp32 = scale.copy(Bitmap.Config.ARGB_8888, true);
    Utils.bitmapToMat(bmp32, origMat);
    Imgproc.cvtColor(origMat, mGray, Imgproc.COLOR_RGB2GRAY);
}

[Lauszus]

Please read the following papers: http://www.face-rec.org/algorithms/PCA/jcn.pdf and https://cseweb.ucsd.edu/classes/wi14/cse152-a/fisherface-pami97.pdf.

And my reports: https://github.com/Lauszus/FaceRecognitionLib/blob/master/Eigenfaces_Report.pdf and https://github.com/Lauszus/FaceRecognitionLib/blob/master/Final_Project_Report.pdf.

[safi099]

I solved the issue of face cropping now it properly train data. but did not recognize the image properly. Actually I do not have understanding of eigen faces and PCA concepts so in order to understand I read the mentioned Papers but did not get my answer. So please do one favour that using Eigen faces what values I can set for these variables. How many train images = ? FaceThreshold values =? DistanceThreshold value =?

Please Answer me. I am very thankful to you. Thanks

[safi099]

??

[Lauszus]

The number of training images are simply limited, as the app became quite slow on old devices. If you have a modern phone, then you can set it quite high.

If there was a magic constant value for the distances that worked for all scenarios, then I would have hard-coded them. You need to experiment with the values yourself to find what suit your needs.