【OpenCV】贝叶斯之肤色分割模型

利用公开肤色数据集训练了一个肤色分类器，OpenCV中有现成的方法，拿来即用。在实际项目中应用，自制数据样本会更有效。

Talk is cheap. Show you the code.

#include <opencv2/core/core.hpp>
#include <opencv2/imgproc/imgproc.hpp>
#include <opencv2/highgui/highgui.hpp>
#include <opencv2/ml/ml.hpp>

#include <fstream>
#include <iostream>

#ifdef _DEBUG
#pragma comment(lib,"opencv_core249d.lib")
#pragma comment(lib,"opencv_highgui249d.lib")
#pragma comment(lib,"opencv_imgproc249d.lib")
#pragma comment(lib,"opencv_ml249d.lib")
#else
#pragma comment(lib,"opencv_core249.lib")
#pragma comment(lib,"opencv_highgui249.lib")
#pragma comment(lib,"opencv_imgproc249.lib")
#pragma comment(lib,"opencv_ml249.lib")
#endif

using namespace std;
using namespace cv;
// split the data & label
bool changeDataType(vector<Vec4i> source, Mat & data, Mat & labels )
{
    int nsample = source.size();

    Mat bgr_data = Mat(nsample,1,CV_8UC3);
    Mat tmp_labels = Mat(nsample,1,CV_32F);

    for (int i = 0; i < nsample; i++ )
    {
        // BGR
        bgr_data.ptr<uchar>(i)[0] = source[i][0];
        bgr_data.ptr<uchar>(i)[1] = source[i][1];
        bgr_data.ptr<uchar>(i)[2] = source[i][2];
        // label
        tmp_labels.ptr<float>(i)[0] = source[i][3];
    }

    Mat ycrcb_data;
    cvtColor(bgr_data,ycrcb_data,CV_BGR2YCrCb);

    Mat tmp_ycrcb_data = ycrcb_data.reshape(1,nsample);
    tmp_ycrcb_data.convertTo(data,CV_32F);

    tmp_labels.copyTo(labels);

    return true;
}
// select crcb channel
bool ycrcb2crcb(Mat& ycrcb, Mat& crcb )
{
    assert(ycrcb.type() == CV_32F);

    if(!crcb.empty())
        crcb.release();
    int n =  ycrcb.rows;
    crcb = Mat(n,2,CV_32F);

    for (int i =0 ; i < n; i++)
    {
        crcb.ptr<float>(i)[0] = ycrcb.ptr<float>(i)[1];
        crcb.ptr<float>(i)[1] = ycrcb.ptr<float>(i)[2];
    }

    return true;
}

int _tmain(int argc, _TCHAR* argv[])
{

    //////////////////////////////////////////////////////////////////////////
    // load train data

    string datafile = "skin_dataset/Skin_NonSkin.txt";

    ifstream fp(datafile);

    if(!fp.is_open()) return -1;

    vector<Vec4i> total_data_labels;

    while(!fp.eof())
    {
        int R = 0 ,G = 0 ,B = 0 ;
        int label = 0;
        fp >> B>>G>>R>> label;
        if(label != 1 && label != 2 )
            continue;
        total_data_labels.push_back(Vec4i(B,G,R,label));

    }

    int nsample = total_data_labels.size();
    cout<<"Load "<<nsample<<" samples: "<<endl;

    //////////////////////////////////////////////////////////////////////////
    // split training data and test samples
    random_shuffle(total_data_labels.begin(),total_data_labels.end());
    int train_n = 0.9*nsample;

    vector<Vec4i> train_data_labels,test_data_labels;
    for (int i = 0 ; i < nsample; i++)
    {
        if(i < train_n)
            train_data_labels.push_back(total_data_labels[i]);
        else
            test_data_labels.push_back(total_data_labels[i]);
    }
    cout<<" split: "<< train_n<<endl;
    cout<<" train samples: "<<train_data_labels.size()<<" "<<test_data_labels.size()<<endl;
    cout<<"======================"<<endl;
    // change data type
    Mat train_data,train_labels;
    changeDataType(train_data_labels,train_data,train_labels);
    Mat train_crcb_data;
    ycrcb2crcb(train_data,train_crcb_data);
    //////////////////////////////////////////////////////////////////////////
    // BayesClassifier
    cout<<"Bayes training: "<<endl;
    NormalBayesClassifier *bayes = new NormalBayesClassifier();
    bayes->train(train_crcb_data,train_labels,Mat(),Mat());
    cout<<"======================"<<endl;
    //////////////////////////////////////////////////////////////////////////
    // test sample predict
    cout<<"Test sample predict: "<<endl;
    Mat test_data,test_labels;
    changeDataType(test_data_labels,test_data,test_labels);

    Mat test_crcb_data;
    ycrcb2crcb(test_data,test_crcb_data);

    Mat result(test_data.rows, test_data.cols, CV_32F);
    bayes->predict(test_crcb_data,&result);

    int err = 0;
    for (int i = 0; i < test_data.rows; i++){
        int trueRes = test_labels.ptr<float>(i)[0];
        float predictRes = result.ptr<float>(i)[0];
        if (fabs(trueRes - predictRes) >= FLT_EPSILON){
            err++;
        }
    }
    cout <<"    error No.:"<< err << " precision: "<< 100-err*1.f / test_data.rows * 100 << "% \n";
    //////////////////////////////////////////////////////////////////////////
    // test image
    cout<<"Test natural image: "<<endl;

    Mat source = imread("image/zdm.jpg",1);
    if(source.empty())
        return -1;
    Mat ycrcb_image;
    cvtColor(source,ycrcb_image,CV_BGR2YCrCb);

    int width = source.cols;
    int height = source.rows;
    Mat ycrcb_image_test_matrix = ycrcb_image.reshape(1,width*height);
    Mat ycrcb_image_test_matrix32f;
    ycrcb_image_test_matrix.convertTo(ycrcb_image_test_matrix32f,CV_32F);

    Mat crcb_image_test_matrix32f;
    ycrcb2crcb(ycrcb_image_test_matrix32f,crcb_image_test_matrix32f);

    double t1 = getTickCount();
    Mat output(ycrcb_image_test_matrix32f.rows, ycrcb_image_test_matrix32f.cols, CV_32F);
    bayes->predict(crcb_image_test_matrix32f,&output);

    Mat mask = Mat(height,width,CV_8U);
    for (int i = 0; i < height; i++)
        for(int j = 0 ; j < width; j++)
        {
            float predictRes = output.ptr<float>(i*width + j)[0];
            if(fabs(predictRes - 1) <= FLT_EPSILON)
                mask.ptr<uchar>(i)[j] = 255;
            else
                mask.ptr<uchar>(i)[j] = 0;
        }

    double t2 = getTickCount();
    cout<<"times: "<<(t2-t1)*1000/getTickFrequency()<<" ms"<<endl;

    //////////////////////////////////////////////////////////////////////////
    // dilate
    Mat element = getStructuringElement(MORPH_CROSS,Size(5,5));
    dilate( mask, mask, element );
    erode(mask,mask,element);

    Mat foot;
    source.copyTo(foot,mask);
    imwrite("zdm_crcb.jpg",mask);
    namedWindow("foot",0);
    imshow("foot",mask);
    waitKey();

    return 0;
}


>>>
Load 245057 samples:
        split: 220551
        train samples: 220551 24506
======================
Bayes training:
======================
Test sample predict:
        error No.:103 precision: 99.5797%
Test natural image:
Input image size: 649824 3
times: 243.396 ms

测试结果：
source:

output:

---

---

---

---

---

---

---

---

---

---

---

---

---

---

---

额，该妹子是年轻时的妮可●基德曼！