2.5.2 基于梯度结构相似度的无参考模糊图像质量评价
这是结构相似度的一个扩展,详细的可以下载论文来进行查看。其算法的步骤如下:
1.先将原始的图像使用11 x 11且方差为1的高斯滤波器进行滤波
2.在接着对模糊后的图像进行半径为4的菱形的Sobel算子的边缘提取
3.将提取后的内容进行膨胀
4.在膨胀图像的情况下判断膨胀模块,并在膨胀模块下计算其GSIM值(相关的膨胀子块可以查看论文,论文不难理解)
5.最后将所有的GSIM值加起来除于膨胀模块的个数,便得到模糊值。
相关代码的如下,有点凌乱,且没有加保护措施:
double DefGSIM(Mat frame)
{
double si1,sj1,si2,sj2,Gx,Gy,temp,sum=0,result,count =0;
int i,j;
int scale = 1;
int delta = 0;
int ddepth = CV_16S;
double k1=0.01;
int L=255;
double c1=pow((k1*L), 2);
Mat gray,Gaussian,sobel,dilate_result;
cvtColor(frame, gray, CV_BGR2GRAY);
IplImage *src = &(IplImage(gray));
uchar *src_data = (uchar*)src->imageData;
GaussianBlur(gray, Gaussian, Size(11, 11), 1);
IplImage *gua = &(IplImage(Gaussian));
uchar *gua_data = (uchar*)src->imageData;
//Sobel operation
/// 创建 grad_x 和 grad_y 矩阵
Mat grad_x, grad_y;
Mat abs_grad_x, abs_grad_y;
/// 求 X方向梯度
//Scharr( src_gray, grad_x, ddepth, 1, 0, scale, delta, BORDER_DEFAULT );
Sobel( Gaussian, grad_x, ddepth, 1, 0, 3, scale, delta, BORDER_DEFAULT );
convertScaleAbs( grad_x, abs_grad_x );
/// 求Y方向梯度
//Scharr( src_gray, grad_y, ddepth, 0, 1, scale, delta, BORDER_DEFAULT );
Sobel( Gaussian, grad_y, ddepth, 0, 1, 3, scale, delta, BORDER_DEFAULT );
convertScaleAbs( grad_y, abs_grad_y );
/// 合并梯度(近似)
addWeighted( abs_grad_x, 0.5, abs_grad_y, 0.5, 0, sobel );
Mat element = getStructuringElement( MORPH_RECT,
Size(4,4 )
);
/// 膨胀操作
dilate( sobel, dilate_result, element );
IplImage *img = &(IplImage(dilate_result));
int height = img->height;
int width = img->width;
int step = img->widthStep/sizeof(uchar);
uchar *data = (uchar*)img->imageData;
Mat gsim = Mat::zeros(height, width, CV_64F);
for (i = 1; i<height-1; i++)
{
for(j = 1; j<width-1; j++)
{
if ((data[i*step+j] == 1) && (data[(i-1)*step+j-1] == 1) && (data[(i-1)*step+j] == 1) && (data[(i+1)*step+j] == 1) && (data[i*step+j-1] == 1) && (data[i*step+j+1] == 1) && (data[(i+1)*step+j-1] == 1) && (data[(i+1)*step+j] == 1) && (data[(i+1)*step+j+1] == 1))
{
si1 = 3*src_data[(i-1)*step+j-1] + 10*src_data[i*step+j-1] + 3*src_data[(i+1)*step+j-1] - 3*src_data[(i-1)*step+j+1] - 10*src_data[i*step+j+1] - 3*src_data[(i+1)*step+j+1];
sj1 = 3*src_data[(i-1)*step+j-1] + 10*src_data[(i-1)*step+j] + 3*src_data[(i-1)*step+j+1] - 3*src_data[(i+1)*step+j-1] - 10*src_data[(i+1)*step+j] - 3*src_data[(i+1)*step+j+1];
Gx = abs(si1) + abs(sj1);
si2 = 3*gua_data[(i-1)*step+j-1] + 10*gua_data[i*step+j-1] + 3*gua_data[(i+1)*step+j-1] - 3*gua_data[(i-1)*step+j+1] - 10*gua_data[i*step+j+1] - 3*gua_data[(i+1)*step+j+1];
sj2 = 3*gua_data[(i-1)*step+j-1] + 10*gua_data[(i-1)*step+j] + 3*gua_data[(i-1)*step+j+1] - 3*gua_data[(i+1)*step+j-1] - 10*gua_data[(i+1)*step+j] - 3*gua_data[(i+1)*step+j+1];
Gy = abs(si2) + abs(sj2);
temp = (2*Gx*Gy + c1)/(Gx*Gx+Gy*Gy+c1);
gsim.at<double>(i,j) = temp ;
count = count + 1;
}
}
}
for (int i = 0; i < height;i++)
{
uchar* data = gsim.ptr<uchar>(i);
for (int j = 0; j < width; j++)
{
sum = data[j] + sum;
}
}
result = sum/count;
return result;
}
该算法的复杂度还是比较低的,且运算的结果也还行。