그레이 영상의 히스토그램 $h(x)$를 두 개의 가우시안 분포($g_1(x)$, $g_2(x)$)의 혼합으로 모델링하여 분리하려고 할 때 기준인 decision boundary 값 (threshold value)을 expectation maximization(EM) 알고리즘을 적용하여 구한다. 

 

E-step: compute responsibility of class 2; (for class 1, 1-γ_i)

 

 

M-step: compute the weighted means (μ1, μ2), variances (σ1, σ2) and mixing probability (π)

 

 

log-likelihood: 

$$\log L = \sum _{i} \log \left[ (1- \pi) \phi_{\theta_1 } (x_i) + \pi \phi_{\theta_2 }(x_i) \right] $$

decision boundary 값은 responsibility = 0.5인  bin 인덱스를 선택하면 된다.

아래 그림의 왼쪽은 히스토그램, 오른쪽은 최대우도 gaussian fitting 결과와 왼쪽 분포의 responsibility($1-\gamma_i$)를 그린 것이다.

void estimGaussParams(std::vector<double>& data, int start, int end, double *mean, double *var) ;

더보기
void estimGaussParams(std::vector<double>& data, int start, int end, double *mean, double *var) {
    double s = 0, sx = 0, sxx = 0;
    for (int i = start; i <= end; i++) {
        s += data[i];
        sx += data[i] * i;
        sxx += data[i] * i * i;
    }
    *mean = sx / s;
    *var = (sxx - sx * sx / s) / s;
};

void initGuess(std::vector<double>& data, double mean[], double var[], double *mixprob);

더보기
void initGuess(std::vector<double>& data, double mean[], double var[], double *mixprob) {
    int start = -1, end = data.size(); 
    // trim null data;
    while (data[++start] <= 0) ;
    while (data[--end] <= 0) ;
    // split given data into two equal size sets;
    int mid = (end + start) / 2;
    // simple mean and variance;
    estimGaussParams(data, start, mid, &mean[0], &var[0]);
    estimGaussParams(data, mid + 1, end, &mean[1], &var[1]);
    // initial guess for mixing probability;
    *mixprob = 0.5; 
};

#define PI (4.0 * atan(1.))

double gaussDist(double x, double mean, double var) ;  

더보기
double gaussDist(double x, double mean, double var) { 
    // N(mean, var);
    double arg = 0.5 * (x - mean) * (x - mean) / var;
    double factor = 1 / sqrt(2.* PI * var);
    return factor * exp(-arg); 
}

double responsibility2(double x, double mean[], double var[], double mixprob) ;   

더보기
double responsibility2(double x, double mean[], double var[], double mixprob) {   
    double a = (1 - mixprob) * gaussDist(x, mean[0], var[0]);
    double b = mixprob * gaussDist(x, mean[1], var[1]);  
    return b / (a + b); 
}

double weightedMeanVar(std::vector<double>& data, std::vector<double> & gamma, double mean[], double var[]) ;

더보기
double weightedMeanVar(std::vector<double>& data, std::vector<double>& gamma, double mean[], double var[]) { 
	// estimate new means;
    double s = 0, sx0 = 0, sx1 = 0, sg = 0;
    for (int i = data.size(); i-- > 0; ) {
        s   += data[i];
        sg  += data[i] * gamma[i]; 
        sx0 += data[i] * i * (1 - gamma[i]);
        sx1 += data[i] * i * gamma[i];
    }
    mean[0] = sx0 / (s - sg);
    mean[1] = sx1 / sg;
    // variances with new mean;
    double sv0 = 0, sv1 = 0;
    for (i = data.size(); i-- > 0; ) {
        sv0 += data[i] * (i - mean[0]) * (i - mean[0]) * (1 - gamma[i]);
        sv1 += data[i] * (i - mean[1]) * (i - mean[1]) * gamma[i];
    }
    var[0] = sv0 / (s - sg);
    var[1] = sv1 / sg;
    // return mixing probability = mixing ratio for class 2;
    return (sg / s);
};
#define EPSILON  1e-6
// Expectation Maximization algorithm applied to Two component Gaussian Mixture Model;
double emTwoCompGMM(std::vector<double>& data) {
    double mean[2], var[2], mixprob;
    std::vector<double> gamma(data.size());     // responsibilities for class 2;
    initGuess(data, mean, var, &mixprob);
    // begin algorithm;
    while (1) {
        // E-step;
        for (int i = data.size(); i-- > 0; ) 
            gamma[i] = responsibility2(i, mean, var, mixprob);
        double old_mixprob = mixprob;
        // M-step;
        mixprob = weightedMeanVar(data, gamma, mean, var);
        TRACE("mixing probability= %f\n", mixprob);
        // check convergence(usually loglikelihood is tested);
        if (fabs(mixprob - old_mixprob) < EPSILON)
            break;
    }
    // estimate decision boundary;
    int k = data.size();
    while (gamma[--k] >= 0.5) ;
    return (2 * k + 1) / 2.; // = average of ;
};

728x90

'Image Recognition' 카테고리의 다른 글

Kuwahara Filter  (2) 2020.12.28
Moving Average을 이용한 Thresholding  (0) 2020.11.26
Union-Find Connected Component Labeling  (0) 2012.11.01
RANSAC: Ellipse Fitting  (1) 2012.10.07
Autofocus Algorithm  (0) 2012.06.03
Posted by helloktk
,