图像分割(Image Segmentation)
作者:王先荣
前言
图像分割指的是将数字图像细分为多个图像子区域的过程,在OpenCv中实现了三种跟图像分割相关的算法,它们分别是:分水岭分割算法、金字塔分割算法以及均值漂移分割算法。它们的使用过程都很简单,下面的文章权且用于记录,并使该系列保持完整吧。
分水岭分割算法
分水岭分割算法需要您或者先前算法提供标记,该标记用于指定哪些大致区域是目标,哪些大致区域是背景等等;分水岭分割算法的分割效果严重依赖于提供的标记。OpenCv中的函数cvWatershed实现了该算法,函数定义如下:
void
cvWatershed(
const
CvArr
*
image, CvArr
*
markers)
其中:image为8为三通道的彩色图像;
markers是单通道整型图像,它用不同的正整数来标记不同的区域,下面的代码演示了如果响应鼠标事件,并生成标记图像。
//
当鼠标按下并在源图像上移动时,在源图像上绘制分割线条
private
void
pbSource_MouseMove(
object
sender, MouseEventArgs e)
{
//
如果按下了左键
if
(e.Button
==
MouseButtons.Left)
{
if
(previousMouseLocation.X
>=
0
&&
previousMouseLocation.Y
>=
0
)
{
Point p1
=
new
Point((
int
)(previousMouseLocation.X
*
xScale), (
int
)(previousMouseLocation.Y
*
yScale));
Point p2
=
new
Point((
int
)(e.Location.X
*
xScale), (
int
)(e.Location.Y
*
yScale));
LineSegment2D ls
=
new
LineSegment2D(p1, p2);
int
thickness
=
(
int
)(LineWidth
*
xScale);
imageSourceClone.Draw(ls,
new
Bgr(255d, 255d, 255d), thickness);
pbSource.Image
=
imageSourceClone.Bitmap;
imageMarkers.Draw(ls,
new
Gray(drawCount), thickness);
}
previousMouseLocation
=
e.Location;
}
}
//
当松开鼠标左键时,将绘图的前一位置设置为(-1,-1)
private
void
pbSource_MouseUp(
object
sender, MouseEventArgs e)
{
previousMouseLocation
=
new
Point(
-
1
,
-
1
);
drawCount
++
;
}
您可以用类似下面的方式来使用分水岭算法:
///
<summary>
///
分水岭算法图像分割
///
</summary>
///
<returns>
返回用时
</returns>
private
string
Watershed()
{
//
分水岭算法分割
Image
<
Gray, Int32
>
imageMarkers2
=
imageMarkers.Copy();
Stopwatch sw
=
new
Stopwatch();
sw.Start();
CvInvoke.cvWatershed(imageSource.Ptr, imageMarkers2.Ptr);
sw.Stop();
//
将分割的结果转换到256级灰度图像
pbResult.Image
=
imageMarkers2.Bitmap;
imageMarkers2.Dispose();
return
string
.Format(
"
分水岭图像分割,用时:{0:F05}毫秒。\r\n
"
, sw.Elapsed.TotalMilliseconds);
}
金字塔分割算法
金字塔分割算法由cvPrySegmentation所实现,该函数的使用很简单;需要注意的是图像的尺寸以及金字塔的层数,图像的宽度和高度必须能被2整除,能够被2整除的次数决定了金字塔的最大层数。下面的代码演示了如果校验金字塔层数:
///
<summary>
///
当改变金字塔分割的参数“金字塔层数”时,对参数进行校验
///
</summary>
///
<param name="sender"></param>
///
<param name="e"></param>
private
void
txtPSLevel_TextChanged(
object
sender, EventArgs e)
{
int
level
=
int
.Parse(txtPSLevel.Text);
if
(level
<
1
||
imageSource.Width
%
(
int
)(Math.Pow(
2
, level
-
1
))
!=
0
||
imageSource.Height
%
(
int
)(Math.Pow(
2
, level
-
1
))
!=
0
)
MessageBox.Show(
this
,
"
注意:您输入的金字塔层数不符合要求,计算结果可能会无效。
"
,
"
金字塔层数错误
"
);
}
使用金字塔分割的示例代码如下:
///
<summary>
///
金字塔分割算法
///
</summary>
///
<returns></returns>
private
string
PrySegmentation()
{
//
准备参数
Image
<
Bgr, Byte
>
imageDest
=
new
Image
<
Bgr,
byte
>
(imageSource.Size);
MemStorage storage
=
new
MemStorage();
IntPtr ptrComp
=
IntPtr.Zero;
int
level
=
int
.Parse(txtPSLevel.Text);
double
threshold1
=
double
.Parse(txtPSThreshold1.Text);
double
threshold2
=
double
.Parse(txtPSThreshold2.Text);
//
金字塔分割
Stopwatch sw
=
new
Stopwatch();
sw.Start();
CvInvoke.cvPyrSegmentation(imageSource.Ptr, imageDest.Ptr, storage.Ptr,
out
ptrComp, level, threshold1, threshold2);
sw.Stop();
//
显示结果
pbResult.Image
=
imageDest.Bitmap;
//
释放资源
imageDest.Dispose();
storage.Dispose();
return
string
.Format(
"
金字塔分割,用时:{0:F05}毫秒。\r\n
"
, sw.Elapsed.TotalMilliseconds);
}
均值漂移分割算法
均值漂移分割算法由cvPryMeanShiftFiltering所实现,均值漂移分割的金字塔层数只能介于[1,7]之间,您可以用类似下面的代码来使用它:
///
<summary>
///
均值漂移分割算法
///
</summary>
///
<returns></returns>
private
string
PryMeanShiftFiltering()
{
//
准备参数
Image
<
Bgr, Byte
>
imageDest
=
new
Image
<
Bgr,
byte
>
(imageSource.Size);
double
spatialRadius
=
double
.Parse(txtPMSFSpatialRadius.Text);
double
colorRadius
=
double
.Parse(txtPMSFColorRadius.Text);
int
maxLevel
=
int
.Parse(txtPMSFNaxLevel.Text);
int
maxIter
=
int
.Parse(txtPMSFMaxIter.Text);
double
epsilon
=
double
.Parse(txtPMSFEpsilon.Text);
MCvTermCriteria termcrit
=
new
MCvTermCriteria(maxIter, epsilon);
//
均值漂移分割
Stopwatch sw
=
new
Stopwatch();
sw.Start();
OpenCvInvoke.cvPyrMeanShiftFiltering(imageSource.Ptr, imageDest.Ptr, spatialRadius, colorRadius, maxLevel, termcrit);
sw.Stop();
//
显示结果
pbResult.Image
=
imageDest.Bitmap;
//
释放资源
imageDest.Dispose();
return
string
.Format(
"
均值漂移分割,用时:{0:F05}毫秒。\r\n
"
, sw.Elapsed.TotalMilliseconds);
}
函数cvPryMeanShiftFiltering在EmguCv中没有实现,我们可以用下面的方式来使用:
//
均值漂移分割
[DllImport(
"
cv200.dll
"
)]
public
static
extern
void
cvPyrMeanShiftFiltering(IntPtr src, IntPtr dst,
double
spatialRadius,
double
colorRadius,
int
max_level, MCvTermCriteria termcrit);
分割效果及性能对比
上述三种分割算法的效果如何呢?下面我们以它们的默认参数,对一幅2272x1704大小的图像进行分割。得到的结果如下所示:

图1 分水岭分割算法(左图白色的线条用于标记区域)

图2 金字塔分割算法

图3 均值漂移分割算法
从上面我们可以看出:
(1)分水岭分割算法的分割效果效果最好,均值漂移分割算法次之,而金字塔分割算法的效果最差;
(2)均值漂移分割算法效率最高,分水岭分割算法接近于均值漂移算法,金字塔分割算法需要很长的时间。
值得注意的是分水岭算法对标记很敏感,需要仔细而认真的绘制。
本文的完整代码如下:
using
System;
using
System.Collections.Generic;
using
System.ComponentModel;
using
System.Data;
using
System.Drawing;
using
System.Linq;
using
System.Text;
using
System.Windows.Forms;
using
System.Diagnostics;
using
System.Runtime.InteropServices;
using
Emgu.CV;
using
Emgu.CV.CvEnum;
using
Emgu.CV.Structure;
using
Emgu.CV.UI;
namespace
ImageProcessLearn
{
public
partial
class
FormImageSegment : Form
{
//
成员变量
private
string
sourceImageFileName
=
"
wky_tms_2272x1704.jpg
"
;
//
源图像文件名
private
Image
<
Bgr, Byte
>
imageSource
=
null
;
//
源图像
private
Image
<
Bgr, Byte
>
imageSourceClone
=
null
;
//
源图像的克隆
private
Image
<
Gray, Int32
>
imageMarkers
=
null
;
//
标记图像
private
double
xScale
=
1d;
//
原始图像与PictureBox在x轴方向上的缩放
private
double
yScale
=
1d;
//
原始图像与PictureBox在y轴方向上的缩放
private
Point previousMouseLocation
=
new
Point(
-
1
,
-
1
);
//
上次绘制线条时,鼠标所处的位置
private
const
int
LineWidth
=
5
;
//
绘制线条的宽度
private
int
drawCount
=
1
;
//
用户绘制的线条数目,用于指定线条的颜色
public
FormImageSegment()
{
InitializeComponent();
}
//
窗体加载时
private
void
FormImageSegment_Load(
object
sender, EventArgs e)
{
//
设置提示
toolTip.SetToolTip(rbWatershed,
"
可以在源图像上用鼠标绘制大致分割区域线条,该线条用于分水岭算法
"
);
toolTip.SetToolTip(txtPSLevel,
"
金字塔层数跟图像尺寸有关,该值只能是图像尺寸被2整除的次数,否则将得出错误结果
"
);
toolTip.SetToolTip(txtPSThreshold1,
"
建立连接的错误阀值
"
);
toolTip.SetToolTip(txtPSThreshold2,
"
分割簇的错误阀值
"
);
toolTip.SetToolTip(txtPMSFSpatialRadius,
"
空间窗的半径
"
);
toolTip.SetToolTip(txtPMSFColorRadius,
"
色彩窗的半径
"
);
toolTip.SetToolTip(btnClearMarkers,
"
清除绘制在源图像上,用于分水岭算法的大致分割区域线条
"
);
//
加载图像
LoadImage();
}
//
当窗体关闭时,释放资源
private
void
FormImageSegment_FormClosing(
object
sender, FormClosingEventArgs e)
{
if
(imageSource
!=
null
)
imageSource.Dispose();
if
(imageSourceClone
!=
null
)
imageSourceClone.Dispose();
if
(imageMarkers
!=
null
)
imageMarkers.Dispose();
}
//
加载源图像
private
void
btnLoadImage_Click(
object
sender, EventArgs e)
{
OpenFileDialog ofd
=
new
OpenFileDialog();
ofd.CheckFileExists
=
true
;
ofd.DefaultExt
=
"
jpg
"
;
ofd.Filter
=
"
图片文件|*.jpg;*.png;*.bmp|所有文件|*.*
"
;
if
(ofd.ShowDialog(
this
)
==
DialogResult.OK)
{
if
(ofd.FileName
!=
""
)
{
sourceImageFileName
=
ofd.FileName;
LoadImage();
}
}
ofd.Dispose();
}
//
清除分割线条
private
void
btnClearMarkers_Click(
object
sender, EventArgs e)
{
if
(imageSourceClone
!=
null
)
imageSourceClone.Dispose();
imageSourceClone
=
imageSource.Copy();
pbSource.Image
=
imageSourceClone.Bitmap;
imageMarkers.SetZero();
drawCount
=
1
;
}
//
当鼠标按下并在源图像上移动时,在源图像上绘制分割线条
private
void
pbSource_MouseMove(
object
sender, MouseEventArgs e)
{
//
如果按下了左键
if
(e.Button
==
MouseButtons.Left)
{
if
(previousMouseLocation.X
>=
0
&&
previousMouseLocation.Y
>=
0
)
{
Point p1
=
new
Point((
int
)(previousMouseLocation.X
*
xScale), (
int
)(previousMouseLocation.Y
*
yScale));
Point p2
=
new
Point((
int
)(e.Location.X
*
xScale), (
int
)(e.Location.Y
*
yScale));
LineSegment2D ls
=
new
LineSegment2D(p1, p2);
int
thickness
=
(
int
)(LineWidth
*
xScale);
imageSourceClone.Draw(ls,
new
Bgr(255d, 255d, 255d), thickness);
pbSource.Image
=
imageSourceClone.Bitmap;
imageMarkers.Draw(ls,
new
Gray(drawCount), thickness);
}
previousMouseLocation
=
e.Location;
}
}
//
当松开鼠标左键时,将绘图的前一位置设置为(-1,-1)
private
void
pbSource_MouseUp(
object
sender, MouseEventArgs e)
{
previousMouseLocation
=
new
Point(
-
1
,
-
1
);
drawCount
++
;
}
//
加载源图像
private
void
LoadImage()
{
if
(imageSource
!=
null
)
imageSource.Dispose();
imageSource
=
new
Image
<
Bgr,
byte
>
(sourceImageFileName);
if
(imageSourceClone
!=
null
)
imageSourceClone.Dispose();
imageSourceClone
=
imageSource.Copy();
pbSource.Image
=
imageSourceClone.Bitmap;
if
(imageMarkers
!=
null
)
imageMarkers.Dispose();
imageMarkers
=
new
Image
<
Gray, Int32
>
(imageSource.Size);
imageMarkers.SetZero();
xScale
=
1d
*
imageSource.Width
/
pbSource.Width;
yScale
=
1d
*
imageSource.Height
/
pbSource.Height;
drawCount
=
1
;
}
//
分割图像
private
void
btnImageSegment_Click(
object
sender, EventArgs e)
{
if
(rbWatershed.Checked)
txtResult.Text
+=
Watershed();
else
if
(rbPrySegmentation.Checked)
txtResult.Text
+=
PrySegmentation();
else
if
(rbPryMeanShiftFiltering.Checked)
txtResult.Text
+=
PryMeanShiftFiltering();
}
///
<summary>
///
分水岭算法图像分割
///
</summary>
///
<returns>
返回用时
</returns>
private
string
Watershed()
{
//
分水岭算法分割
Image
<
Gray, Int32
>
imageMarkers2
=
imageMarkers.Copy();
Stopwatch sw
=
new
Stopwatch();
sw.Start();
CvInvoke.cvWatershed(imageSource.Ptr, imageMarkers2.Ptr);
sw.Stop();
//
将分割的结果转换到256级灰度图像
pbResult.Image
=
imageMarkers2.Bitmap;
imageMarkers2.Dispose();
return
string
.Format(
"
分水岭图像分割,用时:{0:F05}毫秒。\r\n
"
, sw.Elapsed.TotalMilliseconds);
}
///
<summary>
///
金字塔分割算法
///
</summary>
///
<returns></returns>
private
string
PrySegmentation()
{
//
准备参数
Image
<
Bgr, Byte
>
imageDest
=
new
Image
<
Bgr,
byte
>
(imageSource.Size);
MemStorage storage
=
new
MemStorage();
IntPtr ptrComp
=
IntPtr.Zero;
int
level
=
int
.Parse(txtPSLevel.Text);
double
threshold1
=
double
.Parse(txtPSThreshold1.Text);
double
threshold2
=
double
.Parse(txtPSThreshold2.Text);
//
金字塔分割
Stopwatch sw
=
new
Stopwatch();
sw.Start();
CvInvoke.cvPyrSegmentation(imageSource.Ptr, imageDest.Ptr, storage.Ptr,
out
ptrComp, level, threshold1, threshold2);
sw.Stop();
//
显示结果
pbResult.Image
=
imageDest.Bitmap;
//
释放资源
imageDest.Dispose();
storage.Dispose();
return
string
.Format(
"
金字塔分割,用时:{0:F05}毫秒。\r\n
"
, sw.Elapsed.TotalMilliseconds);
}
///
<summary>
///
均值漂移分割算法
///
</summary>
///
<returns></returns>
private
string
PryMeanShiftFiltering()
{
//
准备参数
Image
<
Bgr, Byte
>
imageDest
=
new
Image
<
Bgr,
byte
>
(imageSource.Size);
double
spatialRadius
=
double
.Parse(txtPMSFSpatialRadius.Text);
double
colorRadius
=
double
.Parse(txtPMSFColorRadius.Text);
int
maxLevel
=
int
.Parse(txtPMSFNaxLevel.Text);
int
maxIter
=
int
.Parse(txtPMSFMaxIter.Text);
double
epsilon
=
double
.Parse(txtPMSFEpsilon.Text);
MCvTermCriteria termcrit
=
new
MCvTermCriteria(maxIter, epsilon);
//
均值漂移分割
Stopwatch sw
=
new
Stopwatch();
sw.Start();
OpenCvInvoke.cvPyrMeanShiftFiltering(imageSource.Ptr, imageDest.Ptr, spatialRadius, colorRadius, maxLevel, termcrit);
sw.Stop();
//
显示结果
pbResult.Image
=
imageDest.Bitmap;
//
释放资源
imageDest.Dispose();
return
string
.Format(
"
均值漂移分割,用时:{0:F05}毫秒。\r\n
"
, sw.Elapsed.TotalMilliseconds);
}
///
<summary>
///
当改变金字塔分割的参数“金字塔层数”时,对参数进行校验
///
</summary>
///
<param name="sender"></param>
///
<param name="e"></param>
private
void
txtPSLevel_TextChanged(
object
sender, EventArgs e)
{
int
level
=
int
.Parse(txtPSLevel.Text);
if
(level
<
1
||
imageSource.Width
%
(
int
)(Math.Pow(
2
, level
-
1
))
!=
0
||
imageSource.Height
%
(
int
)(Math.Pow(
2
, level
-
1
))
!=
0
)
MessageBox.Show(
this
,
"
注意:您输入的金字塔层数不符合要求,计算结果可能会无效。
"
,
"
金字塔层数错误
"
);
}
///
<summary>
///
当改变均值漂移分割的参数“金字塔层数”时,对参数进行校验
///
</summary>
///
<param name="sender"></param>
///
<param name="e"></param>
private
void
txtPMSFNaxLevel_TextChanged(
object
sender, EventArgs e)
{
int
maxLevel
=
int
.Parse(txtPMSFNaxLevel.Text);
if
(maxLevel
<
0
||
maxLevel
>
8
)
MessageBox.Show(
this
,
"
注意:均值漂移分割的金字塔层数只能在0至8之间。
"
,
"
金字塔层数错误
"
);
}
}
}

685

被折叠的 条评论
为什么被折叠?



