from: OpenCV单kinect多帧静止场景的深度图像去噪

老板kinect去噪的任务下达已经有半个多月了，前期除了看了几天文献之外就打酱油了，好像每天都很忙，可是就是不知道在忙什么。这几天为了交差，就胡乱凑了几段代码，得到一个结果，也知道不行，先应付一下，再图打算。

程序思想很简单，先对静止的场景连续采样若干帧，然后对所有点在时间域取中值，对取完中值之后的无效点在空间域取最近邻，勉强将黑窟窿填上了。由于代码较长，现在奉上关键的几个片段：

#include<cv.h>
#include<highgui.h>
#include<iostream>
using namespace std;
#ifndef _DENOISE
#define _DENOISE
const int nFrames = 9; // number of consecutive frames
const int width = 640; // frame width
const int height = 480; // frame height
class kinectDenoising
{
private:
IplImage* denoisedImage;
IplImage* frameSet[nFrames];
unsigned int numOfFrames;
CvRect imageROI;
public:
kinectDenoising();
~kinectDenoising();
void addFrame(IplImage* img);
void setImageROI(bool isUpdate = true);
void medianFiltering();
void nearestFiltering();
void updateFrameSet(IplImage* img);
void showDenoiedImage(const char* window);
void showCurrentImage(const char* window);
};
void insertSort(unsigned short* data,int& len,unsigned short newData);
#endif

这是定义的头文件，装模作样的写了一个类，在构造函数里面，除了对denoisedImage分配内存之外其他都置0，析构函数需要释放denoisedImage和frameSet数组的内存。numOfFrames本来设计为frameSet中的图像的帧数，结果由于偷懒就用了一个定长的数组。

[cpp] view plain copy

void kinectDenoising::setImageROI(bool isUpdate)
{
if(!isUpdate)
{
imageROI = cvRect(22,44,591,434);
}
else
{
IplImage* image8u = cvCreateImage(cvSize(width,height),IPL_DEPTH_8U,1);
IplImage* bitImage = cvCreateImage(cvSize(width,height),IPL_DEPTH_8U,1);
// cvThreshold can only handle images of 8UC1 or 32FC1
cvConvertScale(frameSet[0],image8u,255.0/4096.0);
cvThreshold(image8u,bitImage,0,1,CV_THRESH_BINARY);
// the two mats rowReduced and colReduced have to be CV_32SC1 type
// for function cvReduce() seems not to suitable for 16U type and
// 8U type doesn't have enough room for the result.
CvMat* rowReduced = cvCreateMat(1,bitImage->width,CV_32FC1);
// bitImage->width represents number of cols, while bitImage->height stands for rows
CvMat* colReduced = cvCreateMat(bitImage->height,1,CV_32FC1);
cvReduce(bitImage,rowReduced,0,CV_REDUCE_SUM);
cvReduce(bitImage,colReduced,1,CV_REDUCE_SUM);
// compute imageROI.x
for(int i=0;i<rowReduced->cols;i++)
{
float temp = CV_MAT_ELEM(*rowReduced,float,0,i);
if(temp>bitImage->height/3)
{
imageROI.x = i;
break;
}
}
// computer imageROI.width
for(int i=rowReduced->cols;i>0;i--)
{
float temp = CV_MAT_ELEM(*rowReduced,float,0,i-1);
if(temp>bitImage->height/3)
{
imageROI.width = i-imageROI.x;
break;
}
}
// compute imageROI.y
for(int i=0;i<colReduced->rows;i++)
{
float temp = CV_MAT_ELEM(*colReduced,float,i,0);
if(temp>bitImage->height/3)
{
imageROI.y = i;
break;
}
}
// compute imageROI.height
for(int i=colReduced->rows;i>0;i--)
{
float temp = CV_MAT_ELEM(*colReduced,float,i-1,0);
if(temp>bitImage->height/3)
{
imageROI.height = i-imageROI.y;
break;
}
}
// set memory free
cvReleaseImage(&bitImage);
cvReleaseImage(&image8u);
cvReleaseMat(&rowReduced);
cvReleaseMat(&colReduced);
}
}

这是计算深度图像的滤波范围。由于深度图像和彩色图像的视点不一致，导致了将深度图像映射到彩色图像上时有效像素会缩小，典型的现象就是在深度图像的四周会出现黑色的区域。这个函数就是用来将四周的黑色框框去掉。用OpenCV的投影的方法。由于cvReduce()函数要进行累积和的计算，为了不使数据溢出，目标数组应该用32位的浮点型（此函数只支持8位unsigned char型和32位float型）。

[cpp] view plain copy

void kinectDenoising::medianFiltering()
{
// set result image zero
cvSetZero(denoisedImage);
unsigned short data[nFrames];
int total;
for(int i=imageROI.y;i<imageROI.y+imageROI.height;i++)
{
unsigned short* denoisedImageData = (unsigned short*)(denoisedImage->imageData+denoisedImage->widthStep*i);
for(int j=imageROI.x;j<imageROI.x+imageROI.width;j++)
{
total = 0;
for(int k=0;k<nFrames;k++)
{
insertSort(data,total,CV_IMAGE_ELEM(frameSet[k],unsigned short,i,j));
}
if(total != 0)
{
denoisedImageData[j] = data[total/2];
}
}
}
}

中值滤波，统计有效点并排序，然后取中值。insertSort()函数用来将值按从小到大的顺序进行插入，鉴于篇幅的关系，就不贴出来了。

[cpp] view plain copy

void kinectDenoising::nearestFiltering()
{
CvPoint topLeft,downRight;
IplImage* tempImage = cvCloneImage(denoisedImage);
for(int i=imageROI.y;i<imageROI.y+imageROI.height;i++)
{
unsigned short* data = (unsigned short*)(denoisedImage->imageData+denoisedImage->widthStep*i);
for(int j=imageROI.x;j<imageROI.x+imageROI.width;j++)
{
for(int k=1;data[j]==0;k++)
{
topLeft = cvPoint(j-k,i-k); // j为行数 i为列数
downRight = cvPoint(j+k,i+k);
for(int m=topLeft.x;(m<=downRight.x) && (data[j]==0);m++)
{
if(m<0) continue;
if(m>=width) break;
if(topLeft.y>=0)
{
unsigned short temp = CV_IMAGE_ELEM(tempImage,unsigned short,topLeft.y,m);
if(temp > 0)
{
data[j] = temp;
break;
}
}
if(downRight.y < height)
{
unsigned short temp = CV_IMAGE_ELEM(tempImage,unsigned short,downRight.y,m);
if(temp > 0)
{
data[j] = temp;
break;
}
}
}
for(int m=topLeft.y;(m<downRight.y) && (data[j]==0);m++)
{
if(m<0) continue;
if(m>=height) break;
if(topLeft.x>0)
{
unsigned short temp = CV_IMAGE_ELEM(tempImage,unsigned short,m,topLeft.x);
if(temp > 0)
{
data[j] = temp;
break;
}
}
if(downRight.x<width)
{
unsigned short temp = CV_IMAGE_ELEM(tempImage,unsigned short,m,downRight.x);
if(temp > 0)
{
data[j] = temp;
break;
}
}
}
}
}
}
cvReleaseImage(&tempImage);
}

最后是中值滤波，从最内层开始，一层层往外扩，直到找到有效值为止。

运行结果：

源图像：

OpenCV单kinect多帧静止场景的深度图像去噪

结果图像：

OpenCV单kinect多帧静止场景的深度图像去噪

附注：本来这个程序是在8位图像上进行的。先取得16位的unsigned short型深度图像，然后通过cvConvertScale()函数将其转化为8位的unsigned char型，结果在进行去噪的时候怎么都不对，将unsigned char型的数据放到matlab中一看，发现在unsigned short型数据中为0值的像素莫名其妙的在unsigned char型里有了一个很小的值（比如说1, 2, 3, 4, 5什么的，就是不为0）。很奇怪，不知道OpenCV中是怎么搞的。看来还是源数据靠谱，于是将其改为16位的unsigned short型，结果形势一片大好。

http://blog.****.net/chenli2010/article/details/7006573