在OpenCV中自适应确定canny算法的分割门限

oskycar 2016-04-21

展开全文

在OpenCV中用canny算子进行边缘检测速度很快，不过有点不爽的就是高低阈值需要输入。在OpenCV中自适应确定canny算法的分割门限一文仿照matlab中的做法，对canny函数进行了修改，以便当用户没有指定高低阈值时，由函数自适应确定阈值。代码如下：

[cpp] view plain copy

print?

// 仿照matlab，自适应求高低两个门限
CV_IMPL void AdaptiveFindThreshold(CvMat *dx, CvMat *dy, double *low, double *high)
{
CvSize size;
IplImage *imge=0;
int i,j;
CvHistogram *hist;
int hist_size = 255;
float range_0[]={0,256};
float* ranges[] = { range_0 };
double PercentOfPixelsNotEdges = 0.7;
size = cvGetSize(dx);
imge = cvCreateImage(size, IPL_DEPTH_32F, 1);
// 计算边缘的强度, 并存于图像中
float maxv = 0;
for(i = 0; i < size.height; i++ )
{
const short* _dx = (short*)(dx->data.ptr + dx->step*i);
const short* _dy = (short*)(dy->data.ptr + dy->step*i);
float* _image = (float *)(imge->imageData + imge->widthStep*i);
for(j = 0; j < size.width; j++)
{
_image[j] = (float)(abs(_dx[j]) + abs(_dy[j]));
maxv = maxv < _image[j] ? _image[j]: maxv;
}
}
// 计算直方图
range_0[1] = maxv;
hist_size = (int)(hist_size > maxv ? maxv:hist_size);
hist = cvCreateHist(1, &hist_size, CV_HIST_ARRAY, ranges, 1);
cvCalcHist( &imge, hist, 0, NULL );
int total = (int)(size.height * size.width * PercentOfPixelsNotEdges);
float sum=0;
int icount = hist->mat.dim[0].size;
float *h = (float*)cvPtr1D( hist->bins, 0 );
for(i = 0; i < icount; i++)
{
sum += h[i];
if( sum > total )
break;
}
// 计算高低门限
*high = (i+1) * maxv / hist_size ;
*low = *high * 0.4;
cvReleaseImage( &imge );
cvReleaseHist(&hist);
}

然后对cvCanny函数进行以下修改。
在函数体中，当程序用两个sobel算子计算完水平和垂直两个方向的梯度强度过后加入以下代码

[cpp] view plain copy

print?

// 自适应确定阈值
if(low_thresh == -1 && high_thresh == -1)
{
AdaptiveFindThreshold(dx, dy, &low_thresh, &high_thresh);
}

这样，在调用cvCanny函数时，指定高低门限为－1，则cvCanny函数就自适应确定门限。

最后，重新编译cv库，对lib和dll库进行更新。

但是上述代码存在一个问题，当图片是全黑的图时，maxv计算的结果为0，在调用cvCanny检测时，会造成段错误。另外，为了免去修改源码以及重新编译cv库，修改代码如下：

[cpp] view plain copy

print?

void AdaptiveFindThreshold(const CvArr* image, double *low, double *high, int aperture_size=3)
{
cv::Mat src = cv::cvarrToMat(image);
const int cn = src.channels();
cv::Mat dx(src.rows, src.cols, CV_16SC(cn));
cv::Mat dy(src.rows, src.cols, CV_16SC(cn));
cv::Sobel(src, dx, CV_16S, 1, 0, aperture_size, 1, 0, cv::BORDER_REPLIC);
cv::Sobel(src, dy, CV_16S, 0, 1, aperture_size, 1, 0, cv::BORDER_REPLIC);
CvMat _dx = dx, _dy = dy;
_AdaptiveFindThreshold(&_dx, &_dy, low, high);
}
// 仿照matlab，自适应求高低两个门限
void _AdaptiveFindThreshold(CvMat *dx, CvMat *dy, double *low, double *high)
{
CvSize size;
IplImage *imge=0;
int i,j;
CvHistogram *hist;
int hist_size = 255;
float range_0[]={0,256};
float* ranges[] = { range_0 };
double PercentOfPixelsNotEdges = 0.7;
size = cvGetSize(dx);
imge = cvCreateImage(size, IPL_DEPTH_32F, 1);
// 计算边缘的强度, 并存于图像中
float maxv = 0;
for(i = 0; i < size.height; i++ )
{
const short* _dx = (short*)(dx->data.ptr + dx->step*i);
const short* _dy = (short*)(dy->data.ptr + dy->step*i);
float* _image = (float *)(imge->imageData + imge->widthStep*i);
for(j = 0; j < size.width; j++)
{
_image[j] = (float)(abs(_dx[j]) + abs(_dy[j]));
maxv = maxv < _image[j] ? _image[j]: maxv;
}
}
if(maxv == 0){
*high = 0;
*low = 0;
cvReleaseImage( &imge );
return;
}
// 计算直方图
range_0[1] = maxv;
hist_size = (int)(hist_size > maxv ? maxv:hist_size);
hist = cvCreateHist(1, &hist_size, CV_HIST_ARRAY, ranges, 1);
cvCalcHist( &imge, hist, 0, NULL );
int total = (int)(size.height * size.width * PercentOfPixelsNotEdges);
float sum=0;
int icount = hist->mat.dim[0].size;
float *h = (float*)cvPtr1D( hist->bins, 0 );
for(i = 0; i < icount; i++)
{
sum += h[i];
if( sum > total )
break;
}
// 计算高低门限
*high = (i+1) * maxv / hist_size ;
*low = *high * 0.4;
cvReleaseImage( &imge );
cvReleaseHist(&hist);
}