MTCNN移植安卓並檢測視頻中人臉

繼續此前的文章，使用vlc播放了rtsp流媒體視頻後，想檢測視頻中的人臉，之前採用了opencv但是遇到低頭、擡頭和側臉時候，效果就不太好。所以本篇介紹如何使用mtcnn來檢測視頻中的人臉。
大致流程：

一、Tensorflow 模型固化

將PNet、ONet、RNet 網絡參數.npy固化成.pb格式，方便java載入， 固化後的文件在assets中，文件名mtcnn_freezed_model.pb。

二、引入android tensorflow lite 庫

只需在build.gradle(module)最後添加以下幾行語句即可。

dependencies {
    implementation fileTree(include: ['*.jar'], dir: 'libs')
    implementation 'com.android.support:appcompat-v7:27.1.1'
    implementation 'com.android.support.constraint:constraint-layout:1.1.3'
    testImplementation 'junit:junit:4.12'
    androidTestImplementation 'com.android.support.test:runner:1.0.2'
    androidTestImplementation 'com.android.support.test.espresso:espresso-core:3.0.2'
    compile(name: 'libvlc-3.0.0', ext: 'aar')
    implementation files('libs/androidutils.jar')
    compile 'org.tensorflow:tensorflow-android:+'
    implementation files('libs/libutils.jar')
}

三、新建MTCNN類

該類包含加載模型文件，並檢測bitmap中的人臉

package com.cayden.face.facenet;

import android.content.res.AssetManager;
import android.graphics.Bitmap;
import android.graphics.Matrix;
import android.graphics.Point;
import android.util.Log;


import com.cayden.face.FaceApplication;

import org.tensorflow.contrib.android.TensorFlowInferenceInterface;

import java.util.Vector;

import static java.lang.Math.max;
import static java.lang.Math.min;

/**
 *  Created by caydencui on 2018/9/6.
 */
public class MTCNN {
    //參數
    private float factor=0.709f;
    private float PNetThreshold=0.6f;
    private float RNetThreshold=0.7f;
    private float ONetThreshold=0.7f;
    //MODEL PATH
    private static final String MODEL_FILE  = "file:///android_asset/mtcnn_freezed_model.pb";
    //tensor name
    private static final String PNetInName  ="pnet/input:0";
    private static final String[] PNetOutName =new String[]{"pnet/prob1:0","pnet/conv4-2/BiasAdd:0"};
    private static final String RNetInName  ="rnet/input:0";
    private static final String[] RNetOutName =new String[]{ "rnet/prob1:0","rnet/conv5-2/conv5-2:0",};
    private static final String ONetInName  ="onet/input:0";
    private static final String[] ONetOutName =new String[]{ "onet/prob1:0","onet/conv6-2/conv6-2:0","onet/conv6-3/conv6-3:0"};


    private static class SingletonInstance {
        private static final MTCNN INSTANCE = new MTCNN();
    }

    public static MTCNN getInstance() {
        return SingletonInstance.INSTANCE;
    }

    //安卓相關
    public  long lastProcessTime;   //最後一張圖片處理的時間ms
    private static final String TAG="MTCNN";
    private AssetManager assetManager;
    private TensorFlowInferenceInterface inferenceInterface;

    private MTCNN() {
        assetManager= FaceApplication.getMyApplication().getAssets();
        loadModel();
    }



    private boolean loadModel() {
        //AssetManager
        try {
            inferenceInterface = new TensorFlowInferenceInterface(assetManager, MODEL_FILE);
            Log.d("MTCNN","[*]load model success");
        }catch(Exception e){
            Log.e("MTCNN","[*]load model failed"+e);
            return false;
        }
        return true;
    }
    //讀取Bitmap像素值，預處理(-127.5 /128)，轉化爲一維數組返回
    private float[] normalizeImage(Bitmap bitmap){
        int w=bitmap.getWidth();
        int h=bitmap.getHeight();
        float[] floatValues=new float[w*h*3];
        int[]   intValues=new int[w*h];
        bitmap.getPixels(intValues,0,bitmap.getWidth(),0,0,bitmap.getWidth(),bitmap.getHeight());
        float imageMean=127.5f;
        float imageStd=128;

        for (int i=0;i<intValues.length;i++){
            final int val=intValues[i];
            floatValues[i * 3 + 0] = (((val >> 16) & 0xFF) - imageMean) / imageStd;
            floatValues[i * 3 + 1] = (((val >> 8) & 0xFF) - imageMean) / imageStd;
            floatValues[i * 3 + 2] = ((val & 0xFF) - imageMean) / imageStd;
        }
        return floatValues;
    }
    /*
       檢測人臉,minSize是最小的人臉像素值
     */
    private Bitmap bitmapResize(Bitmap bm, float scale) {
        int width = bm.getWidth();
        int height = bm.getHeight();
        // CREATE A MATRIX FOR THE MANIPULATION。matrix指定圖片仿射變換參數
        Matrix matrix = new Matrix();
        // RESIZE THE BIT MAP
        matrix.postScale(scale, scale);
        Bitmap resizedBitmap = Bitmap.createBitmap(
                bm, 0, 0, width, height, matrix, true);
        return resizedBitmap;
    }
    //輸入前要翻轉，輸出也要翻轉
    private  int PNetForward(Bitmap bitmap, float [][]PNetOutProb, float[][][]PNetOutBias){
        int w=bitmap.getWidth();
        int h=bitmap.getHeight();

        float[] PNetIn=normalizeImage(bitmap);
        Utils.flip_diag(PNetIn,h,w,3); //沿着對角線翻轉
        inferenceInterface.feed(PNetInName,PNetIn,1,w,h,3);
        inferenceInterface.run(PNetOutName,false);
        int PNetOutSizeW=(int) Math.ceil(w*0.5-5);
        int PNetOutSizeH=(int) Math.ceil(h*0.5-5);
        float[] PNetOutP=new float[PNetOutSizeW*PNetOutSizeH*2];
        float[] PNetOutB=new float[PNetOutSizeW*PNetOutSizeH*4];
        inferenceInterface.fetch(PNetOutName[0],PNetOutP);
        inferenceInterface.fetch(PNetOutName[1],PNetOutB);
        //【寫法一】先翻轉，後轉爲2/3維數組
        Utils.flip_diag(PNetOutP,PNetOutSizeW,PNetOutSizeH,2);
        Utils.flip_diag(PNetOutB,PNetOutSizeW,PNetOutSizeH,4);
        Utils.expand(PNetOutB,PNetOutBias);
        Utils.expandProb(PNetOutP,PNetOutProb);
        /*
        *【寫法二】這個比較快，快了3ms。意義不大，用上面的方法比較直觀
        for (int y=0;y<PNetOutSizeH;y++)
            for (int x=0;x<PNetOutSizeW;x++){
               int idx=PNetOutSizeH*x+y;
               PNetOutProb[y][x]=PNetOutP[idx*2+1];
               for(int i=0;i<4;i++)
                   PNetOutBias[y][x][i]=PNetOutB[idx*4+i];
            }
        */
        return 0;
    }
    //Non-Maximum Suppression
    //nms，不符合條件的deleted設置爲true
    private void nms(Vector<Box> boxes, float threshold, String method){
        //NMS.兩兩比對
        //int delete_cnt=0;
        for(int i=0;i<boxes.size();i++) {
            Box box = boxes.get(i);
            if (!box.deleted) {
                //score<0表示當前矩形框被刪除
                for (int j = i + 1; j < boxes.size(); j++) {
                    Box box2=boxes.get(j);
                    if (!box2.deleted) {
                        int x1 = max(box.box[0], box2.box[0]);
                        int y1 = max(box.box[1], box2.box[1]);
                        int x2 = min(box.box[2], box2.box[2]);
                        int y2 = min(box.box[3], box2.box[3]);
                        if (x2 < x1 || y2 < y1) continue;
                        int areaIoU = (x2 - x1 + 1) * (y2 - y1 + 1);
                        float iou=0f;
                        if (method.equals("Union"))
                            iou = 1.0f*areaIoU / (box.area() + box2.area() - areaIoU);
                        else if (method.equals("Min"))
                            iou= 1.0f*areaIoU / (min(box.area(),box2.area()));
                        if (iou >= threshold) { //刪除prob小的那個框
                            if (box.score>box2.score)
                                box2.deleted=true;
                            else
                                box.deleted=true;
                            //delete_cnt++;
                        }
                    }
                }
            }
        }
        //Log.i(TAG,"[*]sum:"+boxes.size()+" delete:"+delete_cnt);
    }
    private int generateBoxes(float[][] prob,float[][][]bias,float scale,float threshold,Vector<Box> boxes){
        int h=prob.length;
        int w=prob[0].length;
        //Log.i(TAG,"[*]height:"+prob.length+" width:"+prob[0].length);
        for (int y=0;y<h;y++)
            for (int x=0;x<w;x++){
                float score=prob[y][x];
                //only accept prob >threadshold(0.6 here)
                if (score>PNetThreshold){
                    Box box=new Box();
                    //score
                    box.score=score;
                    //box
                    box.box[0]= Math.round(x*2/scale);
                    box.box[1]= Math.round(y*2/scale);
                    box.box[2]= Math.round((x*2+11)/scale);
                    box.box[3]= Math.round((y*2+11)/scale);
                    //bbr
                    for(int i=0;i<4;i++)
                        box.bbr[i]=bias[y][x][i];
                    //add
                    boxes.addElement(box);
                }
            }
        return 0;
    }
    private void BoundingBoxReggression(Vector<Box> boxes){
        for (int i=0;i<boxes.size();i++)
            boxes.get(i).calibrate();
    }
    //Pnet + Bounding Box Regression + Non-Maximum Regression
    /* NMS執行完後，才執行Regression
     * (1) For each scale , use NMS with threshold=0.5
     * (2) For all candidates , use NMS with threshold=0.7
     * (3) Calibrate Bounding Box
     * 注意：CNN輸入圖片最上面一行，座標爲[0..width,0]。所以Bitmap需要對摺後再跑網絡;網絡輸出同理.
     */
    private Vector<Box> PNet(Bitmap bitmap, int minSize){
        int whMin=min(bitmap.getWidth(),bitmap.getHeight());
        float currentFaceSize=minSize;  //currentFaceSize=minSize/(factor^k) k=0,1,2... until excced whMin
        Vector<Box> totalBoxes=new Vector<Box>();
        //【1】Image Paramid and Feed to Pnet
        while (currentFaceSize<=whMin){
            float scale=12.0f/currentFaceSize;
            //(1)Image Resize
            Bitmap bm=bitmapResize(bitmap,scale);
            int w=bm.getWidth();
            int h=bm.getHeight();
            //(2)RUN CNN
            int PNetOutSizeW=(int)(Math.ceil(w*0.5-5)+0.5);
            int PNetOutSizeH=(int)(Math.ceil(h*0.5-5)+0.5);
            float[][]   PNetOutProb=new float[PNetOutSizeH][PNetOutSizeW];;
            float[][][] PNetOutBias=new float[PNetOutSizeH][PNetOutSizeW][4];
            PNetForward(bm,PNetOutProb,PNetOutBias);
            //(3)數據解析
            Vector<Box> curBoxes=new Vector<Box>();
            generateBoxes(PNetOutProb,PNetOutBias,scale,PNetThreshold,curBoxes);
            //Log.i(TAG,"[*]CNN Output Box number:"+curBoxes.size()+" Scale:"+scale);
            //(4)nms 0.5
            nms(curBoxes,0.5f,"Union");
            //(5)add to totalBoxes
            for (int i=0;i<curBoxes.size();i++)
                if (!curBoxes.get(i).deleted)
                    totalBoxes.addElement(curBoxes.get(i));
            //Face Size等比遞增
            currentFaceSize/=factor;
        }
        //NMS 0.7
        nms(totalBoxes,0.7f,"Union");
        //BBR
        BoundingBoxReggression(totalBoxes);
        return Utils.updateBoxes(totalBoxes);
    }
    //截取box中指定的矩形框(越界要處理)，並resize到size*size大小，返回數據存放到data中。
    public Bitmap tmp_bm;
    private void crop_and_resize(Bitmap bitmap, Box box, int size, float[] data){
        //(2)crop and resize
        Matrix matrix = new Matrix();
        float scale=1.0f*size/box.width();
        matrix.postScale(scale, scale);
        Bitmap croped= Bitmap.createBitmap(bitmap, box.left(),box.top(),box.width(), box.height(),matrix,true);
        //(3)save
        int[] pixels_buf=new int[size*size];
        croped.getPixels(pixels_buf,0,croped.getWidth(),0,0,croped.getWidth(),croped.getHeight());
        float imageMean=127.5f;
        float imageStd=128;
        for (int i=0;i<pixels_buf.length;i++){
            final int val=pixels_buf[i];
            data[i * 3 + 0] = (((val >> 16) & 0xFF) - imageMean) / imageStd;
            data[i * 3 + 1] = (((val >> 8) & 0xFF) - imageMean) / imageStd;
            data[i * 3 + 2] = ((val & 0xFF) - imageMean) / imageStd;
        }
    }
    /*
     * RNET跑神經網絡，將score和bias寫入boxes
     */
    private void RNetForward(float[] RNetIn,Vector<Box> boxes){
        int num=RNetIn.length/24/24/3;
        //feed & run
        inferenceInterface.feed(RNetInName,RNetIn,num,24,24,3);
        inferenceInterface.run(RNetOutName,false);
        //fetch
        float[] RNetP=new float[num*2];
        float[] RNetB=new float[num*4];
        inferenceInterface.fetch(RNetOutName[0],RNetP);
        inferenceInterface.fetch(RNetOutName[1],RNetB);
        //轉換
        for (int i=0;i<num;i++) {
            boxes.get(i).score = RNetP[i * 2 + 1];
            for (int j=0;j<4;j++)
                boxes.get(i).bbr[j]=RNetB[i*4+j];
        }
    }
    //Refine Net
    private Vector<Box> RNet(Bitmap bitmap, Vector<Box> boxes){
        //RNet Input Init
        int num=boxes.size();
        float[] RNetIn=new float[num*24*24*3];
        float[] curCrop=new float[24*24*3];
        int RNetInIdx=0;
        for (int i=0;i<num;i++){
            crop_and_resize(bitmap,boxes.get(i),24,curCrop);
            Utils.flip_diag(curCrop,24,24,3);
            //Log.i(TAG,"[*]Pixels values:"+curCrop[0]+" "+curCrop[1]);
            for (int j=0;j<curCrop.length;j++) RNetIn[RNetInIdx++]= curCrop[j];
        }
        //Run RNet
        RNetForward(RNetIn,boxes);
        //RNetThreshold
        for (int i=0;i<num;i++)
            if (boxes.get(i).score<RNetThreshold)
                boxes.get(i).deleted=true;
        //Nms
        nms(boxes,0.7f,"Union");
        BoundingBoxReggression(boxes);
        return Utils.updateBoxes(boxes);
    }
    /*
     * ONet跑神經網絡，將score和bias寫入boxes
     */
    private void ONetForward(float[] ONetIn,Vector<Box> boxes){
        int num=ONetIn.length/48/48/3;
        //feed & run
        inferenceInterface.feed(ONetInName,ONetIn,num,48,48,3);
        inferenceInterface.run(ONetOutName,false);
        //fetch
        float[] ONetP=new float[num*2]; //prob
        float[] ONetB=new float[num*4]; //bias
        float[] ONetL=new float[num*10]; //landmark
        inferenceInterface.fetch(ONetOutName[0],ONetP);
        inferenceInterface.fetch(ONetOutName[1],ONetB);
        inferenceInterface.fetch(ONetOutName[2],ONetL);
        //轉換
        for (int i=0;i<num;i++) {
            //prob
            boxes.get(i).score = ONetP[i * 2 + 1];
            //bias
            for (int j=0;j<4;j++)
                boxes.get(i).bbr[j]=ONetB[i*4+j];

            //landmark
            for (int j=0;j<5;j++) {
                int x=boxes.get(i).left()+(int) (ONetL[i * 10 + j]*boxes.get(i).width());
                int y= boxes.get(i).top()+(int) (ONetL[i * 10 + j +5]*boxes.get(i).height());
                boxes.get(i).landmark[j] = new Point(x,y);
                //Log.i(TAG,"[*] landmarkd "+x+ "  "+y);
            }
        }
    }
    //ONet
    private Vector<Box> ONet(Bitmap bitmap, Vector<Box> boxes){
        //ONet Input Init
        int num=boxes.size();
        float[] ONetIn=new float[num*48*48*3];
        float[] curCrop=new float[48*48*3];
        int ONetInIdx=0;
        for (int i=0;i<num;i++){
            crop_and_resize(bitmap,boxes.get(i),48,curCrop);
            Utils.flip_diag(curCrop,48,48,3);
            for (int j=0;j<curCrop.length;j++) ONetIn[ONetInIdx++]= curCrop[j];
        }
        //Run ONet
        ONetForward(ONetIn,boxes);
        //ONetThreshold
        for (int i=0;i<num;i++)
            if (boxes.get(i).score<ONetThreshold)
                boxes.get(i).deleted=true;
        BoundingBoxReggression(boxes);
        //Nms
        nms(boxes,0.7f,"Min");
        return Utils.updateBoxes(boxes);
    }
    private void square_limit(Vector<Box> boxes, int w, int h){
        //square
        for (int i=0;i<boxes.size();i++) {
            boxes.get(i).toSquareShape();
            boxes.get(i).limit_square(w,h);
        }
    }
    /*
     * 參數：
     *   bitmap:要處理的圖片
     *   minFaceSize:最小的人臉像素值.(此值越大，檢測越快)
     * 返回：
     *   人臉框
     */
    public Vector<Box> detectFaces(Bitmap bitmap, int minFaceSize) {
        long t_start = System.currentTimeMillis();
        //【1】PNet generate candidate boxes
        Vector<Box> boxes=PNet(bitmap,minFaceSize);
        square_limit(boxes,bitmap.getWidth(),bitmap.getHeight());
        //【2】RNet
        boxes=RNet(bitmap,boxes);
        square_limit(boxes,bitmap.getWidth(),bitmap.getHeight());
        //【3】ONet
        boxes=ONet(bitmap,boxes);
        //return
        Log.i(TAG,"[*]Mtcnn Detection Time:"+(System.currentTimeMillis()-t_start));
        lastProcessTime=(System.currentTimeMillis()-t_start);
        return  boxes;
    }
}

四、視頻流數據處理

由於vlc播放的流媒體視頻格式是nv12，需要將其轉爲nv21，並保存爲bitmap

• 1、首先給出nv12轉爲nv21方法

    private void NV12ToNV21(byte[] nv12, byte[] nv21, int width, int height) {
        if (nv21 == null || nv12 == null) return;
        int framesize = width * height;
        int i = 0, j = 0;
        //System.arraycopy(nv21, test, nv12, test, framesize);
        for (i = 0; i < framesize; i++) {
            nv21[i] = nv12[i];
        }
        for (j = 0; j < framesize / 2; j += 2) {
            nv21[framesize + j] = nv12[j + framesize + 1];
        }
        for (j = 0; j < framesize / 2; j += 2) {
            nv21[framesize + j + 1] = nv12[j + framesize];
        }
    }

• 2、然後給出nv21高效率轉換爲bitmap的方法
  可以查看此前的一篇文章 nv21高效率轉換爲bitmap
• 3、調用mtcnn的人臉檢測方法

Vector boxes = mtcnn.detectFaces(bitmap, 20);

• 4、根據返回的數據 ，標記人臉

    protected void drawAnim(Vector<Box> faces, SurfaceView outputView, float scale_bit, int cameraId, String fps) {
        Paint paint = new Paint();
        Canvas canvas = ((SurfaceView) outputView).getHolder().lockCanvas();
        if (canvas != null) {
            try {
                int viewH = outputView.getHeight();
                int viewW = outputView.getWidth();
//                DLog.d("viewW:"+viewW+",viewH:"+viewH);
                canvas.drawColor(0, PorterDuff.Mode.CLEAR);
                if (faces == null || faces.size() == 0) return;
                for (int i = 0; i < faces.size(); i++) {
                    paint.setColor(Color.BLUE);
                    int size = DisplayUtil.dip2px(this, 3);
                    paint.setStrokeWidth(size);
                    paint.setStyle(Paint.Style.STROKE);
                    Box box = faces.get(i);
                    float[] rect = box.transform2float();
                    float x1 = rect[0] * scale_bit;
                    float y1 = rect[1] * scale_bit;
                    float rect_width = rect[2] * 0.5F;
                    RectF rectf = new RectF(x1, y1, x1 + rect_width, y1 + rect_width);
                    canvas.drawRect(rectf, paint);
                }

            } catch (Exception e) {
                e.printStackTrace();
            } finally {
                ((SurfaceView) outputView).getHolder().unlockCanvasAndPost(canvas);
            }
        }
    }

項目源碼

https://github.com/cayden/facesample
本項目主要基於vlc來播放流媒體視頻

主要包含以下內容

• 1、使用已經編譯的libvlc來播放流媒體視頻
• 2、使用MTCNN進行人臉識別並標記人臉
• 3、保存標記的人臉圖片
• 4、使用FACENET進行人臉比對
• 未完待續…

v1.0.2

• 1, 通過MTCNN檢測人臉
• 2, 對人臉進行標記

v1.0.1

• 1, 獲取返回的視頻流數據
• 2, 將數據nv12轉換爲nv21,並保存圖片

v1.0.0

• 1, added libvlc
• 2, support for playing rtsp video stream

感謝大家的閱讀，也希望能轉發並關注我的公衆號