人脸识别，顾名思义，就是通过人脸对比的方式，得出人脸相识度的过程。区别于人脸检测。
对于OpenCV的人脸检测，实现流程，请看我之前写的博客：
OpenCV导入
OpenCV人脸检测
OpenCV竖屏检测
本次人脸识别，实现思路如下：
（一）读取本地数据源作为对比凭证源
（二）动态读取视频捕获的人脸数据，于对比凭证源进行对比
开始发车：
（一）读取本地数据源作为对比凭证
本次做法，为了方便演示，首先，准备了一些100px*100px像素的数据源图片，放到了指定的目录，然后通过File类的listFile()操作，在使用前，把目录下的的文件放到到内存中，这里使用OpenCV的bitmap转mat的方法。具体代码实现如下图：
/**
* 初始化
*/
private void init() {
try {
File path = new File(FACE_FILE_PATH);
path.mkdirs();
File[] allFile = path.listFiles();
for (File cache : allFile) {
Bitmap cacheBitmap = BitmapFactory.decodeFile(cache.getPath());
Mat mat1 = new Mat();
Utils.bitmapToMat(cacheBitmap, mat1);
Mat result = new Mat();
Imgproc.cvtColor(mat1, result, Imgproc.COLOR_BGR2GRAY);
Moments mom = Imgproc.moments(result, false);
putLocal(mom.toString(), mat1);
}
} catch (Exception e) {
}
}
上图代码，就把本地所有的数据集合，放到了内存中。
（二）动态读取视频捕获的人脸数据
通过前面章节，已经可以动态监测到人脸数据了，这里直接使用即可，代码如下图：
@Override
public Mat onCameraFrame(Mat aInputFrame) {
Imgproc.cvtColor(aInputFrame, grayscaleImage, Imgproc.COLOR_RGBA2RGB);
MatOfRect faces = new MatOfRect();
if (cascadeClassifier != null) {
cascadeClassifier.detectMultiScale(grayscaleImage, faces, 1.1, 3, 2,
new Size(absoluteFaceSize, absoluteFaceSize), new Size());
}
Rect[] facesArray = faces.toArray();
for (int i = 0; i < facesArray.length; i++) {
detectFace(aInputFrame, facesArray[i]);
Imgproc.rectangle(aInputFrame, facesArray[i].tl(), facesArray[i].br(), new Scalar(0, 255, 0, 255), 3);
}
return aInputFrame;
}
主要是调用detectFace()方法，把人脸帧和人脸数据，传递给外部进行相关处理。下图为监测人脸后的具体实现代码：
@Override
public void detectFace(final Mat source, final Rect face) {
super.detectFace(source, face);
ThreadManager.getImgExecutors().execute(new Runnable() {
@Override
public void run() {
try {
Mat sub = source.submat(face);
Mat mat = new Mat();
Size size = new Size(100, 100);
Imgproc.resize(sub, mat, size);
Mat result = new Mat();
Imgproc.cvtColor(mat, result, Imgproc.COLOR_BGR2GRAY);
Moments mom = Imgproc.moments(result, false);
FaceDetectManager.getInstance().putFace(mom.m00 + "", mat);
FaceDetectManager.getInstance().check();
} catch (Exception e) {
e.printStackTrace();
Log.d("error2", "error set:" + e.getMessage());
}
}
});
}
可以看到也是保存了裁剪后的人脸图片，像素为100px*100px。
然后就调用FaceDetectManager的相关方法。这里要注意的是，开了一个线程，为什么要开线程，请自行脑补。
最后放上对比的核心代码：
/**
* 检测特征值
*/
public void check() {
if (System.currentTimeMillis() - mCheckTime > 2000) {
mCheckTime = System.currentTimeMillis();
} else {
return;
}
//开始检测
ThreadManager.getImgExecutors().execute(new Runnable() {
@Override
public void run() {
Log.d("check", "mFaceLocalMap大小：" + mFaceLocalMap.size() + "\tmFaceMemoryMap大小：" + mFaceMemoryMap.size());
for (Map.Entry<String, Mat> local : mFaceLocalMap.entrySet()) {
for (Map.Entry<String, Mat> memory : mFaceMemoryMap.entrySet()) {
Mat src1 = new Mat();
Imgproc.cvtColor(local.getValue(), src1, Imgproc.COLOR_BGR2GRAY);
Mat target1 = new Mat();
Imgproc.cvtColor(memory.getValue(), target1, Imgproc.COLOR_BGR2GRAY);
src1.convertTo(src1, CvType.CV_32F);
target1.convertTo(target1, CvType.CV_32F);
double similar = Imgproc.compareHist(src1, target1, Imgproc.CV_COMP_CORREL);
Log.e("相识度", "相似度 ： ==" + similar);
src1.release();
target1.release();
if (similar > 0.6) {
if (mListener != null) {
mListener.result(similar, local.getValue());
}
}
}
}
mFaceMemoryMap.clear();
}
});
}
主要还是通过Imgproc.compareHist()方法对mat之间进行对比，最后得出相似度。然后回调出去。
注意：mat与bitmap之间的转换，如果使用了OpenCV的转换函数的话，记得分辨率保持一致，而且mat要求传入的是RGB格式的数据源，而bitmap的格式要求则要是ARGB_8888或者RGB_565。实现如下图：
FaceDetectManager.getInstance().setSimilarResultListener(new FaceDetectManager.SimilarResultListener() {
@Override
public void result(final double similar, final Mat result) {
runOnUiThread(new Runnable() {
@Override
public void run() {
try {
if (similar > 0.8) {
Bitmap bitmap = Bitmap.createBitmap(100, 100, Bitmap.Config.RGB_565);
Utils.matToBitmap(result, bitmap);
getTargetView().setImageBitmap(bitmap);
}
} catch (Exception e) {
Log.d("error", "error set:" + e.getMessage() + "\t\t" + (getTargetView() == null));
}
}
});
}
});
OpenCV转换源码如下图：
/**
* Converts OpenCV Mat to Android Bitmap.
* <p>
* <br>This function converts an image in the OpenCV Mat representation to the Android Bitmap.
* <br>The input Mat object has to be of the types 'CV_8UC1' (gray-scale), 'CV_8UC3' (RGB) or 'CV_8UC4' (RGBA).
* <br>The output Bitmap object has to be of the same size as the input Mat and of the types 'ARGB_8888' or 'RGB_565'.
* <br>This function throws an exception if the conversion fails.
*
* @param mat is a valid input Mat object of types 'CV_8UC1', 'CV_8UC3' or 'CV_8UC4'.
* @param bmp is a valid Bitmap object of the same size as the Mat and of type 'ARGB_8888' or 'RGB_565'.
* @param premultiplyAlpha is a flag, that determines, whether the Mat needs to be converted to alpha premultiplied format (like Android keeps 'ARGB_8888' bitmaps); the flag is ignored for 'RGB_565' bitmaps.
*/
public static void matToBitmap(Mat mat, Bitmap bmp, boolean premultiplyAlpha) {
if (mat == null)
throw new IllegalArgumentException("mat == null");
if (bmp == null)
throw new IllegalArgumentException("bmp == null");
nMatToBitmap2(mat.nativeObj, bmp, premultiplyAlpha);
}
/**
* Short form of the <b>matToBitmap(mat, bmp, premultiplyAlpha=false)</b>
* @param mat is a valid input Mat object of the types 'CV_8UC1', 'CV_8UC3' or 'CV_8UC4'.
* @param bmp is a valid Bitmap object of the same size as the Mat and of type 'ARGB_8888' or 'RGB_565'.
*/
public static void matToBitmap(Mat mat, Bitmap bmp) {
matToBitmap(mat, bmp, false);
}
相关要求请看英文描述。
至此，完成，是不是非常简单。
that's all------------------------------------------------------------------------------
网友评论