Skimage Hog

11 Apr 2018

SVM+Hog在一些图像算法中也还是有一定的存在价值，在将scikit-image的hog特征计算用c语言实现时，c计算的结果和python计算的结果总是有一点偏差。

float orie_flag[10] = {0.0, 0.363969900364, 0.8390986263, 1.73204726945, 5.67124270781,
            -5.67133071011, -1.73205788383, -0.839103148256, -0.363972905488, -0.0};
Mat sobel_y = Mat::zeros(Size(64,64), CV_32FC1);
Mat sobel_x = Mat::zeros(Size(64,64), CV_32FC1);
Mat magn = Mat::zeros(Size(64,64), CV_32FC1);
Mat orie = Mat::zeros(Size(64,64), CV_32FC1);
float ceils[8*8*9] = {0};
float blocks[7*7*9*4] = {0};

for (int i=0; i<64; i++)
{
    for (int j=0; j<64; j++)
    {
        gray.at<float>(i,j) = sqrt(gray.at<float>(i,j));
    }
}

for (int i=0; i<64; i++)
{
    for (int j=1; j<63; j++)
    {
        sobel_x.at<float>(i, j) = gray.at<float>(i,j+1) - gray.at<float>(i,j-1);
    }
}

for (int i=1; i<63; i++)
{
    for (int j=0; j<64; j++)
    {
        sobel_y.at<float>(i, j) = gray.at<float>(i+1,j) - gray.at<float>(i-1,j);
    }
}

for (int i=0; i<64; i++)
{
    for (int j=0; j<64; j++)
    {
        magn.at<float>(i, j) = sqrt(sobel_x.at<float>(i, j)*sobel_x.at<float>(i, j) + sobel_y.at<float>(i, j)*sobel_y.at<float>(i, j));
    }
}

for (int i=0; i<64; i++)
{
    for (int j=0; j<64; j++)
    {
        orie.at<float>(i, j) = sobel_y.at<float>(i, j) / (sobel_x.at<float>(i, j) + 0.00001);
    }
}

for (int i=0; i<8; i++)
{
    for (int j=0; j<8; j++)
    {
        for (int m=0; m<8; m++)
        {
            for (int n=0; n<8; n++)
            {
                int k;
                float tmp = orie.at<float>(i*8+m, j*8+n);
                if (tmp >= 0)
                {
                    for (k=0; k<4; k++)
                    {
                        if (tmp >= orie_flag[k] && tmp < orie_flag[k+1])
                        {
                            break;
                        }
                    }
                    ceils[(i*8+j)*9 + k] += magn.at<float>(i*8+m, j*8+n);
                }
                else
                {
                    for (k=9; k>5; k--)
                    {
                        if (tmp <= orie_flag[k] && tmp > orie_flag[k-1])
                        {
                            break;
                        }
                    }
                    k = k-1;
                    ceils[(i*8+j)*9 + k] += magn.at<float>(i*8+m, j*8+n);
                }
            }
        }
    }
}

for (int i=0; i<7; i++)
{
    for (int j=0; j<7; j++)
    {
        float sum_ = 0.00001;
        for (int k=0; k<9; k++)
        {
            sum_ += abs(ceils[(i*8+j)*9 + k]);
            sum_ += abs(ceils[(i*8+(j+1))*9 + k]);
            sum_ += abs(ceils[((i+1)*8+(j+1))*9 + k]);
            sum_ += abs(ceils[((i+1)*8+j)*9 + k]);
        }

        for (int k=0; k<9; k++)
        {
            blocks[(i*7+j)*36 + k] = ceils[(i*8+j)*9 + k] / sum_;
            blocks[(i*7+j)*36 + k + 9] = ceils[(i*8+(j+1))*9 + k] / sum_;
            blocks[(i*7+j)*36 + k + 18] = ceils[((i+1)*8+j)*9 + k] / sum_;
            blocks[(i*7+j)*36 + k + 27] = ceils[((i+1)*8+(j+1))*9 + k] / sum_;
        }
    }
}

for (int i=0; i<7*7*4*9; i++)
{
    features[feat_idx++] = blocks[i];
}

from skimage.feature import hog
hog_feat = hog(img, orientations=9,
                   pixels_per_cell=(8, 8),
                   cells_per_block=(2, 2),
                   transform_sqrt=True,
                   visualise=False, feature_vector=True)

原因：

python中安装的scikit-image是0.13版本的，它在计算hog特征时是这样计算sobel的：

gy, gx = [np.ascontiguousarray(g, dtype=np.double)
              for g in np.gradient(image)]

最新版和scikit-image和0.13之前的版本，在计算hog特征时是这样计算sobel的：

g_row = np.empty(channel.shape, dtype=np.double)
g_row[0, :] = 0
g_row[-1, :] = 0
g_row[1:-1, :] = channel[2:, :] - channel[:-2, :]
g_col = np.empty(channel.shape, dtype=np.double)
g_col[:, 0] = 0
g_col[:, -1] = 0
g_col[:, 1:-1] = channel[:, 2:] - channel[:, :-2]

在C代码中它的区别就应该是这样的：

for (int i=0; i<64; i++)
{
    sobel_x.at<float>(i, 0) = (gray.at<float>(i,1) - gray.at<float>(i,0));
    for (int j=1; j<63; j++)
    {
        sobel_x.at<float>(i, j) = (gray.at<float>(i,j+1) - gray.at<float>(i,j-1)) / 2;
    }
    sobel_x.at<float>(i, 63) = (gray.at<float>(i,63) - gray.at<float>(i,62));
}

for (int j=0; j<64; j++)
{
    sobel_y.at<float>(0, j) = (gray.at<float>(1,j) - gray.at<float>(0,j));
    for (int i=1; i<63; i++)
    {
        sobel_y.at<float>(i, j) = (gray.at<float>(i+1,j) - gray.at<float>(i-1,j)) / 2;
    }
    sobel_y.at<float>(63, j) = (gray.at<float>(63,j) - gray.at<float>(62,j));
}

for (int i=0; i<64; i++)
{
    for (int j=1; j<63; j++)
    {
        sobel_x.at<float>(i, j) = gray.at<float>(i,j+1) - gray.at<float>(i,j-1);
    }
}

for (int i=1; i<63; i++)
{
    for (int j=0; j<64; j++)
    {
        sobel_y.at<float>(i, j) = gray.at<float>(i+1,j) - gray.at<float>(i-1,j);
    }
}

来看下差异吧，都取同一幅图像的第一个Block的Hog特征，它有4个Cell，9个方向，差异还是有一点点的：

[[[ 0.04225977  0.03862688  0.01705302  0.02437312  0.07293652  0.04062629
    0.00249895  0.01547415  0.02562522]
  [ 0.00381061  0.01124655  0.00704325  0.02541224  0.13330107  0.01689826
    0.01232815  0.00643004  0.00499948]]

 [[ 0.00222884  0.01873355  0.02347761  0.04217004  0.04436155  0.02916896
    0.03580889  0.01976478  0.01399942]
  [ 0.00922168  0.01071582  0.01380053  0.10009506  0.09301306  0.01187018
    0.01550533  0.00866581  0.00643561]]]

[[[ 0.05323969  0.040701    0.01576395  0.02282596  0.07068123  0.03564167
    0.00263313  0.00939887  0.02091938]
  [ 0.01391507  0.01185045  0.00475391  0.01755162  0.1328273   0.01264098
    0.01092408  0.00677531  0.00526794]]

 [[ 0.00234852  0.00705782  0.02134361  0.04443442  0.06453884  0.02467325
    0.03773169  0.02082607  0.0049391 ]
  [ 0.00971685  0.01129122  0.01454157  0.1054698   0.09800752  0.01250757
    0.01633791  0.00913113  0.00678118]]]