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 _DENOISEconst int nFrames = 9;   // number of consecutive framesconst int width = 640;   // frame widthconst int height = 480;  // frame heightclass 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。

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 32FC1cvConvertScale(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 freecvReleaseImage(&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）。

void kinectDenoising::medianFiltering(){// set result image zerocvSetZero(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。

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：

结果图像：

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.csdn.net/chenli2010/article/details/7006573