使用liner、feather、multiband对已经拼接的数据进行融合

所谓"blend",英文解释为“vt. 混合vi. 混合；协调n. 混合；掺合物”这里应该理解为是图像数据的融合。这是“识别->对准->融合”的最后一步。融合是决定拼接质量的关键一步，一方面它决定于图像对准的质量，一方面它本身的也直接对拼接的最终结果负责。

     最简单和便于理解的融合为liner，正好借这个例子来说明说明是融合，简单的说，就是在融合的区域……（这个地方引用相关资料）liner在opencv中没有实现，但是本身简单有效，对于要求不是很高的情况可以使用，这里给出函数（再做相关解释）

     #pragma region mulitStitch

/*----------------------------

 * 功|能 : 多图匹配

 *----------------------------

 * 函数y : MulitMatch

 * 访问 : private

 * 返回 : void

 *

 * 参数y : matinput      [in]     全部需要a匹配图的vector

 * 参数y : matloc1       [ot]     所有D匹配中D对应|于第一图的结果向量

 * 参数y : matloc1       [ot]     所有D匹配中D对应|于第二t图的结果向量

 * 参数y : match_method  [in]     匹配方法

 */

void MulitMatch(deque<Mat>& matinput,deque<Point>& matloc1,deque<Point>& matloc2, int match_method)

{

      Mat img_display1;Mat img_display2;

                  Point matchLoc1;Point matchLoc2;

                 for (int i =0;i<matinput.size()-1;i++)

                 {

 

                                 //拷贝副本

                                  img_display1 = matinput[i];

                                  img_display2 = matinput[i+1];

                                 //以中D心区域为aroi

                                 // Mat imagetmp (img_display1, Rect(img_display1.rows/2, img_display1.cols/2, 10, 10) );//11:44:02

                                  

                                 Mat imagetmp (img_display1, Rect(960,240, 10, 10) );//11:44:02

                                 int result_cols =  img_display1.cols - imagetmp.cols + 1;

                                 int result_rows = img_display1.rows - imagetmp.rows + 1;

                                 Mat imagematch;

                                 imagematch.create( result_cols, result_rows, CV_32FC1 );

                                 /// 进行D匹配和标准化

                                 //匹配1

                                 matchTemplate( img_display1, imagetmp, imagematch, match_method );

                                 normalize( imagematch, imagematch, 0, 1, NORM_MINMAX, -1, Mat() );

                                 double minVal; double maxVal;

                                 Point minLoc; Point maxLoc;

                                 minMaxLoc( imagematch, &minVal, &maxVal, &minLoc, &maxLoc, Mat() );

                                 //智能判D断,这a里的matchLoc就是最佳匹配点

                                 if( match_method  == CV_TM_SQDIFF || match_method == CV_TM_SQDIFF_NORMED )

                                 { matchLoc1 = minLoc; }

                                 else

                                 { matchLoc1 = maxLoc; }

                                matloc1.push_back(matchLoc1); //加入序列D

                                 //匹配2

                                 matchTemplate( img_display2, imagetmp, imagematch, match_method );

                                 normalize( imagematch, imagematch, 0, 1, NORM_MINMAX, -1, Mat() );

                                 minMaxLoc( imagematch, &minVal, &maxVal, &minLoc, &maxLoc, Mat() );

                                 //智能判D断,这a里的matchLoc就是最佳匹配点

                                 if( match_method  == CV_TM_SQDIFF || match_method == CV_TM_SQDIFF_NORMED )

                                 { matchLoc2 = minLoc; }

                                 else

                                 { matchLoc2 = maxLoc; }

                                 matloc2.push_back(matchLoc2); //加入序列D

                 }

}

/*----------------------------

 * 功|能 : 多图对准

 *----------------------------

 * 函数y : MulitAlign

 * 访问 : private

 * 返回 : Mat

 *

 * 参数y : matinput      [in]     全部需要a匹配图的vector

 * 参数y : matloc1       [ot]     所有D匹配中D对应|于第一图的结果向量

 * 参数y : matloc1       [ot]     所有D匹配中D对应|于第二t图的结果向量

 */

Mat MulitAlign(deque<Mat>& matinput,deque<Point>& matloc1,deque<Point>& matloc2)

{              

                Mat outImage; //待y输出图片

                 //计算图片大小   

                 int nr = matinput[0].rows;

                 int nl = matinput[0].cols*matinput[0].channels();

                 int ioffset = 0;int ioffsetdetail = 0;

                 //计算offset

                 for (int i =0;i<matloc1.size()-1;i++)

                {

                                ioffset = ioffset+matloc1[i].y- matloc2[i].y;

                }

                outImage.create( matinput[0].rows+ioffset, matinput[0].cols, matinput[0].type());

                 for (int i=0;i<matloc1.size()-1;i++)

                {

                                 if (i == 0)//如果第一弹

                                {

                                                 for (int a=0;a<nr;a++) //第一图

                                                {

                                                                 const uchar* inData=matinput[0].ptr<uchar>(a);

                                                                uchar* outData=outImage.ptr<uchar>(a); 

                                                                 for(int j=0;j<nl;j++)

                                                                {

                                                                                outData[j]=inData[j];           

                                                                }

                                                }

 

                                                 for (int b=0;b<nr;b++) //第二t图

                                                {

                                                                 const uchar* inData=matinput[1].ptr<uchar>(b);

                                                                uchar* outData=outImage.ptr<uchar>(b+matloc1[0].y-matloc2[0].y); 

                                                                 for(int j=0;j<nl;j++)

                                                                {

                                                                                outData[j]=inData[j];           

                                                                }

                                                }

                                                ioffsetdetail += matloc1[0].y-matloc2[0].y;

                                }

                                 else//如果不是第一弹

                                {

                                                 for (int b=0;b<nr;b++) 

                                                {

                                                                 const uchar* inData=matinput[i+1].ptr<uchar>(b);

                                                                uchar* outData=outImage.ptr<uchar>(b+ioffsetdetail+matloc1[i].y-matloc2[i].y); 

                                                                 for(int j=0;j<nl;j++)

                                                                {

                                                                                outData[j]=inData[j];           

                                                                }

                                                }

                                                ioffsetdetail += matloc1[i+1].y-matloc2[i].y;

                                }              

                                                

                }

 

                 return outImage;

}

/*----------------------------

 * 功|能 : 多图融合

 *----------------------------

 * 函数y : MulitBlend

 * 访问 : private

 * 返回 : Mat&

 *

 * 参数y : matinput      [in]     图片输入序列D

 * 参数y : imagesrc      [in]     已经-对准的图片

 * 参数y : matloc1                            [in]     第一图匹配位置

 * 参数y : matloc2        [in]     第二t图匹配位置

 */

Mat MulitBlend(deque<Mat>& matinput, const Mat& imagesrc,deque<Point>& matloc1,deque<Point>& matloc2)

{              

                Mat outImage; //待y输出图片

                imagesrc.copyTo(outImage); //图像拷贝

                 int ioffsetdetail = 0;

                 double dblend = 0.0;

                 for (int i =0;i<matloc1.size()-1;i++)

                {

                                dblend = 0.0;

                                 int ioffset = matloc1[i].y - matloc2[i].y;//row的偏移

                                 for (int j = 0;j<100;j++)//这a个地方用i 和 j很不好

                                {              

                                                outImage.row(ioffsetdetail+ioffset+j) = matinput[i].row(ioffset+j)*(1-dblend)+ matinput[i+1].row(j)*(dblend);

                                                dblend = dblend +0.01;

                                }

                                                ioffsetdetail += ioffset;

                }

                 return outImage;

}

#pragma endregion mulitStitch

 

在最新版的opencv中(至少在2.4.5之后），提供了mulitband和feather函数。jsxyhelu认为，总体来说，mulitband是目前最好的融合算法，（paper),在（2007）这篇经典的图像拼接论文中得到引用，需要提及的一点是opencv的stitch函数主要就是基于2007这篇论文实现的，它的算法实现的第一篇引用论文就是2007。当然，好的算法可能用起来比较麻烦，简单的算法也有其适合使用的地方。

。。。。。。pipleline

mulitblend的主要思想是小频率事件大空间划分，大频率事件小空间划分，具体内容参考论文。featherblend就是常见的所谓“羽化”，这里主要考虑工程实现。

 

首先参考image blander函数

detail::Blender
class detail::Blender
Base class for all blenders.
class CV_EXPORTS Blender
{
public:
virtual ~Blender() {}
enum { NO, FEATHER, MULTI_blend };
static Ptr<Blender> createDefault(int type, bool try_gpu = false);
void prepare(const std::vector<Point> &corners, const std::vector<Size> &sizes);
virtual void prepare(Rect dst_roi);
virtual void feed(const Mat &img, const Mat &mask, Point tl);
virtual void blend(Mat &dst, Mat &dst_mask);
protected:
Mat dst_, dst_mask_;
Rect dst_roi_;
};

 

在detail空间中存在blender函数，是所有blender的基函数。可以实现的包括feather和multiblend两种方法。

 

 blender = Blender::createDefault(blend_type, try_gpu);是创建函数，两个参数决定了采用哪一种blender方法，和是否采用gpu

 

 

detail::Blender::prepare

为blend准备相关数据
C++: void detail::Blender::prepare(const std::vector<Point>& corners, const std::vector<Size>&
sizes)
Parameters
corners – 原始图像文件的左上角点
sizes – 原始文件大小

注意这里的两点都是vector

 

detail::Blender::feed
处理图像
C++: void detail::Blender::feed(const Mat& img, const Mat& mask, Point tl)
Parameters
img –原始图像
mask – mask

tl topleft点

注意这里是一张一张处理图像的

 

detail::Blender::blend
处理blend操作,是最后最后输出操作 ，从blend结构体中返回pano全景图像了
C++: void detail::Blender::blend(Mat& dst, Mat& dst_mask)
Parameters
dst – Final pano
dst_mask – Final pano mask

 

detail::MultiBandBlender
class detail::MultiBandBlender : public detail::Blender
Blender which uses multi-band blending algorithm (see [BA83],就是前面提到的那篇论文).
class CV_EXPORTS MultiBandBlender : public Blender
{
public:
MultiBandBlender(int try_gpu = false, int num_bands = 5);
int numBands() const { return actual_num_bands_; }
void setNumBands(int val) { actual_num_bands_ = val; }

void prepare(Rect dst_roi);
void feed(const Mat &img, const Mat &mask, Point tl);
void blend(Mat &dst, Mat &dst_mask);
private:
/* hidden */
};
See also:
detail::Blender

 

detail::FeatherBlender
class detail::FeatherBlender : public detail::Blender
Simple blender which mixes images at its borders.
class CV_EXPORTS FeatherBlender : public Blender
{
public:
FeatherBlender(float sharpness = 0.02f) { setSharpness(sharpness); }
float sharpness() const { return sharpness_; }
void setSharpness(float val) { sharpness_ = val; }
void prepare(Rect dst_roi);
void feed(const Mat &img, const Mat &mask, Point tl);
void blend(Mat &dst, Mat &dst_mask);
// Creates weight maps for fixed set of source images by their masks and top-left corners.
// Final image can be obtained by simple weighting of the source images.
Rect createWeightMaps(const std::vector<Mat> &masks, const std::vector<Point> &corners,
std::vector<Mat> &weight_maps);
private:
/* hidden */
};
See also:
detail::Blender

这两个函数在文档中都没有详细的解释。这里进行补充说明。

基于cookbook第9章的estimateH.cpp继续前进

将其cv::Mat image1= cv::imread( "parliament1.bmp",1);

                cv::Mat image2= cv::imread( "parliament2.bmp",1);

结果没有问题，拼接的出来了，但是缝合线也非常明显

 

将不需要的代码注释掉，需要注意的是，书中的代码认为图片是从右向左移动的

//需要a注意a的一点是，原-始文件t的图片是按照从右至左边进行D移动的。

                cv::Point* p1 = new cv::Point(image1.cols,1);

                cv::Point* p2 = new cv::Point(image1.cols,image2.rows-1);

                cv::line(result,*p1,*p2,cv::Scalar(255,255,255),2);

                cv::namedWindow( "After warping0");

                cv::imshow( "After warping0",result);

画出一条白线，基本标注位置

 

 

cv::Mat result;

                cv::warpPerspective(image1, // input image

                                result,                                      // output image

                                homography,                         // homography

                                cv::Size(2*image1.cols,image1.rows)); // size of output image

                cv::Mat resultback;

                result.copyTo(resultback);

 

                 // Copy image 1 on the first half of full image

                cv::Mat half(result,cv::Rect(0,0,image2.cols,image2.rows));

                image2.copyTo(half);

    // Display the warp image

                cv::namedWindow( "After warping");

                cv::imshow( "After warping",result);

                 //需要a注意a的一点是，原-始文件t的图片是按照从右至左边进行D移动的。

//             cv::Point* p1 = new cv::Point(image1.cols,1);

//             cv::Point* p2 = new cv::Point(image1.cols,image2.rows-1);

//             cv::line(result,*p1,*p2,cv::Scalar(255,255,255),2);

//             cv::namedWindow("After warping0");

//             cv::imshow("After warping0",result);

                 //进行Dliner的融合

                Mat outImage; //待y输出图片

                result.copyTo(outImage); //图像拷贝

                 double dblend = 0.0;

                 int ioffset =image2.cols-100;//col的初始定位

                 for (int i = 0;i<100;i++)

                {              

                                outImage.col(ioffset+i) = image2.col(ioffset+i)*(1-dblend) + resultback.col(ioffset+i)*dblend;

                                dblend = dblend +0.01;

                }

需要注意的是这里不是将image1和image2进行融合，而是将原始的result和image进行融合。

由于背景比较单一，而且图片分辨率不是很高，所以这个结果融合的 结果非常不