stitching.cpp魚眼影象拼接融合原始碼分析

之前執行OpenCV官方示例的cpp時看到stitching.cpp拼接融合還不錯然後我在MATLAB上用之前編的經緯對映法校正三幅魚眼影象後不知道該怎樣儲存下校正好的圖如果save或者save as 那麼會有figure的白色邊緣不能用來拼接所以我直接截圖儲存為jpg 這樣就沒有白色邊緣了：

校正後的：然後把這三幅MATLAB執行出來的結果傳遞給OpenCV的stitching.cpp 然後執行出來：

現在開始分析影象拼接融合的原始碼：stitching.cpp：

#include <iostream>
#include <fstream>
#include "opencv2/highgui/highgui.hpp"
#include "opencv2/stitching/stitcher.hpp"
using namespace std;
using namespace cv;
bool try_use_gpu = false;
vector<Mat> imgs;
string result_name = "result3.jpg"; //儲存全景圖的檔名可以在工程目錄下找到
void printUsage();
int parseCmdArgs(int argc, char** argv);

int main(int argc, char* argv[])
{
int retval = parseCmdArgs(argc, argv);
if (retval) return -1;

Mat pano;
Stitcher stitcher = Stitcher::createDefault(try_use_gpu); //這個函式預設不使用GPU！
Stitcher::Status status = stitcher.stitch(imgs, pano); //找尋待拼接的兩幅圖的旋轉角度並生成最後的全景圖這個函式才是要分析的重頭戲！
if (status != Stitcher::OK) //確保拼接成功
{
cout << "Can't stitch images, error code = " << int(status) << endl;
return -1;
}
namedWindow("stitching result");
imshow("stitching result", pano);
waitKey(0);
imwrite(result_name,pano);
return 0;
}
void printUsage() //預設不用GPU
{
cout <<
"Rotation model images stitcher.\n\n"
"stitching img1 img2 [...imgN]\n\n"
"Flags:\n"
" --try_use_gpu (yes|no)\n"
" Try to use GPU. The default value is 'no'. All default values\n"
" are for CPU mode.\n"
" --output <result_img>\n"
" The default is 'result.jpg'.\n";
}
int parseCmdArgs(int argc, char** argv) //這個函式就是讀入待拼接的圖片放在imgs這個裝待拼圖片的容器裡面
{
if (argc == 1)
{
printUsage();
return -1;
}
for (int i = 1; i < argc; ++i)
{
if (string(argv[i]) == "--help" || string(argv[i]) == "/?")
{
printUsage();
return -1;
}
else if (string(argv[i]) == "--try_use_gpu")
{
if (string(argv[i + 1]) == "no")
try_use_gpu = false;
else if (string(argv[i + 1]) == "yes")
try_use_gpu = true;
else
{
cout << "Bad --try_use_gpu flag value\n";
return -1;
}
i++;
}
else if (string(argv[i]) == "--output")
{
result_name = argv[i + 1];
i++;
}
else
{
Mat img = imread(argv[i]);
if (img.empty())
{
cout << "Can't read image '" << argv[i] << "'\n";
return -1;
}
imgs.push_back(img);
}
}
return 0;
}

上面stitching.cpp很好理解現在來看涉及影象拼接、融合的函式原始碼：Stitcher::Status status = stitcher.stitch(imgs, pano); 用CMake開啟原始碼檢視原始碼：

可以看到stitcher類下各個函式的定義 stitch()是可過載的而程式裡用的是第一個形式這個定義非常簡單所以接下來又需要去看Status status = estimateTransform(images, rois);（估算幾幅待拼圖之間的關係比如仿射變換旋轉平移縮放等等）和composePanorama(pano);（影象融合組成全景圖）這兩個函式的定義 estimateTransform也簡短

Stitcher::Status Stitcher::estimateTransform(InputArray images, const vector<vector<Rect> > &rois)
{
images.getMatVector(imgs_);
rois_ = rois;
Status status;
if ((status = matchImages()) != OK) //要看這個的原始碼
return status;

//接下來還要看這個函式estimateCameraParams();(估算相機引數)的原始碼
estimateCameraParams();
return OK;
}

於是又接著找到：

Stitcher::Status Stitcher::matchImages()
{
if ((int)imgs_.size() < 2)
{
LOGLN("Need more images");
return ERR_NEED_MORE_IMGS; //待拼影象不能少於2幅圖
}
work_scale_ = 1;
seam_work_aspect_ = 1;
seam_scale_ = 1;
bool is_work_scale_set = false;
bool is_seam_scale_set = false;
Mat full_img, img;
features_.resize(imgs_.size()); //這是？？
seam_est_imgs_.resize(imgs_.size());
full_img_sizes_.resize(imgs_.size());

LOGLN("Finding features...");
#if ENABLE_LOG
int64 t = getTickCount(); //找尋特徵點計時
#endif

for (size_t i = 0; i < imgs_.size(); ++i)
{
full_img = imgs_[i]; //依次讀取每幅待拼影象給full_img
full_img_sizes_[i] = full_img.size(); //將每次讀入影象的大小mxn給full_img_sizes
//registr_resol_是影象匹配的解析度大小，影象的面積尺寸變為registr_resol_*100000 ？？
if (registr_resol_ < 0)
{
img = full_img;
work_scale_ = 1;
is_work_scale_set = true;
}
else
{
if (!is_work_scale_set)
{

//預處理將影象縮放full_img.size().area()表示面積m、n的乘積

//計算work_scale，將影象resize到面積在registr_resol_*10^6以下
work_scale_ = min(1.0, sqrt(registr_resol_ * 1e6 / full_img.size().area()));
is_work_scale_set = true;
}

//將full_img和img的尺寸都縮放到計算出來的work_scale_下？！
resize(full_img, img, Size(), work_scale_, work_scale_);
}
if (!is_seam_scale_set)
{

//seam_est_resol_是拼接縫畫素的大小，和匹配預設值一樣有預設值？
seam_scale_ = min(1.0, sqrt(seam_est_resol_ * 1e6 / full_img.size().area()));
seam_work_aspect_ = seam_scale_ / work_scale_; //和上個if中的差不多意思
is_seam_scale_set = true;
}

if (rois_.empty())

   //如果rois_是空矩陣就找尋特徵點給features_ （這個待會看原始碼）
(*features_finder_)(img, features_[i]);
else
{
vector<Rect> rois(rois_[i].size());
for (size_t j = 0; j < rois_[i].size(); ++j)
{
Point tl(cvRound(rois_[i][j].x * work_scale_), cvRound(rois_[i][j].y * work_scale_));
Point br(cvRound(rois_[i][j].br().x * work_scale_), cvRound(rois_[i][j].br().y * work_scale_));
rois[j] = Rect(tl, br);
}
(*features_finder_)(img, features_[i], rois);
}
features_[i].img_idx = (int)i; //說明是第幾幅圖的特徵點
LOGLN("Features in image #" << i+1 << ": " << features_[i].keypoints.size());
  //將源影象resize到seam_scale_*10^6，並存入seam_est_imgs_[]中
resize(full_img, img, Size(), seam_scale_, seam_scale_);
seam_est_imgs_[i] = img.clone();
}

// Do it to save memory
features_finder_->collectGarbage(); //這裡要看原始碼！！怎麼找尋的
full_img.release(); //release釋放待會兒去讀下一幅圖
img.release();
//找尋特徵點至此結束計算出時間
LOGLN("Finding features, time: " << ((getTickCount() - t) / getTickFrequency()) << " sec");

LOG("Pairwise matching");
#if ENABLE_LOG
t = getTickCount(); //開始進行匹配match
#endif
(*features_matcher_)(features_, pairwise_matches_, matching_mask_);
features_matcher_->collectGarbage(); //要看原始碼！！
LOGLN("Pairwise matching, time: " << ((getTickCount() - t) / getTickFrequency()) << " sec"); //匹配結束

// Leave only images we are sure are from the same panorama

//conf_thresh_是兩幅圖來自同一全景圖的置信度來判斷讀入的圖片是否屬於同一全景圖
indices_ = detail::leaveBiggestComponent(features_, pairwise_matches_, (float)conf_thresh_);
vector<Mat> seam_est_imgs_subset;
vector<Mat> imgs_subset;
vector<Size> full_img_sizes_subset;
for (size_t i = 0; i < indices_.size(); ++i)
{
imgs_subset.push_back(imgs_[indices_[i]]);
seam_est_imgs_subset.push_back(seam_est_imgs_[indices_[i]]);
full_img_sizes_subset.push_back(full_img_sizes_[indices_[i]]);
}
seam_est_imgs_ = seam_est_imgs_subset;
imgs_ = imgs_subset;
full_img_sizes_ = full_img_sizes_subset;
//檢查由上述篩選來自同一副全景圖的圖片的數量是否大於2幅圖
if ((int)imgs_.size() < 2)
{
LOGLN("Need more images");
return ERR_NEED_MORE_IMGS;
}

return OK;
}

結合這個人的http://blog.csdn.net/hanshuning/article/details/41960401和這個人的http://www.geekcome.com/content-10-8390-1.html分析上面的

在matchImage()分析完後要分析estimateCameraParams()：

void Stitcher::estimateCameraParams()
{
detail::HomographyBasedEstimator estimator;
estimator(features_, pairwise_matches_, cameras_);

for (size_t i = 0; i < cameras_.size(); ++i)
{
Mat R;
cameras_[i].R.convertTo(R, CV_32F);
cameras_[i].R = R;
LOGLN("Initial intrinsic parameters #" << indices_[i] + 1 << ":\n " << cameras_[i].K());
}

bundle_adjuster_->setConfThresh(conf_thresh_);
(*bundle_adjuster_)(features_, pairwise_matches_, cameras_);

// Find median focal length and use it as final image scale
vector<double> focals;
for (size_t i = 0; i < cameras_.size(); ++i)
{
LOGLN("Camera #" << indices_[i] + 1 << ":\n" << cameras_[i].K());
focals.push_back(cameras_[i].focal);
}

std::sort(focals.begin(), focals.end());
if (focals.size() % 2 == 1)
warped_image_scale_ = static_cast<float>(focals[focals.size() / 2]);
else
warped_image_scale_ = static_cast<float>(focals[focals.size() / 2 - 1] + focals[focals.size() / 2]) * 0.5f;

if (do_wave_correct_)
{
vector<Mat> rmats;
for (size_t i = 0; i < cameras_.size(); ++i)
rmats.push_back(cameras_[i].R);
detail::waveCorrect(rmats, wave_correct_kind_);
for (size_t i = 0; i < cameras_.size(); ++i)
cameras_[i].R = rmats[i];
}
}

接下來生成全景圖（影象融合等）：composePanorama(pano);即看這個函式的原始碼

Stitcher::Status Stitcher::composePanorama(OutputArray pano)
{
return composePanorama(vector<Mat>(), pano); //就這一句繼續追蹤下去檢視composePanorama(vector<Mat>(), pano);的原始碼
}
如下所示：
Stitcher::Status Stitcher::composePanorama(InputArray images, OutputArray pano)
{
LOGLN("Warping images (auxiliary)... ");

vector<Mat> imgs;
images.getMatVector(imgs);
if (!imgs.empty())
{
CV_Assert(imgs.size() == imgs_.size());

Mat img;
seam_est_imgs_.resize(imgs.size());

for (size_t i = 0; i < imgs.size(); ++i)
{
imgs_[i] = imgs[i];
resize(imgs[i], img, Size(), seam_scale_, seam_scale_);
seam_est_imgs_[i] = img.clone();
}

vector<Mat> seam_est_imgs_subset;
vector<Mat> imgs_subset;

for (size_t i = 0; i < indices_.size(); ++i)
{
imgs_subset.push_back(imgs_[indices_[i]]);
seam_est_imgs_subset.push_back(seam_est_imgs_[indices_[i]]);
}

seam_est_imgs_ = seam_est_imgs_subset;
imgs_ = imgs_subset;
}

Mat &pano_ = pano.getMatRef();

#if ENABLE_LOG
int64 t = getTickCount();
#endif

vector<Point> corners(imgs_.size());
vector<Mat> masks_warped(imgs_.size());
vector<Mat> images_warped(imgs_.size());
vector<Size> sizes(imgs_.size());
vector<Mat> masks(imgs_.size());

// Prepare image masks
for (size_t i = 0; i < imgs_.size(); ++i)
{
masks[i].create(seam_est_imgs_[i].size(), CV_8U);
masks[i].setTo(Scalar::all(255));
}

// Warp images and their masks
Ptr<detail::RotationWarper> w = warper_->create(float(warped_image_scale_ * seam_work_aspect_));
for (size_t i = 0; i < imgs_.size(); ++i)
{
Mat_<float> K;
cameras_[i].K().convertTo(K, CV_32F);
K(0,0) *= (float)seam_work_aspect_;
K(0,2) *= (float)seam_work_aspect_;
K(1,1) *= (float)seam_work_aspect_;
K(1,2) *= (float)seam_work_aspect_;

corners[i] = w->warp(seam_est_imgs_[i], K, cameras_[i].R, INTER_LINEAR, BORDER_REFLECT, images_warped[i]);
sizes[i] = images_warped[i].size();

w->warp(masks[i], K, cameras_[i].R, INTER_NEAREST, BORDER_CONSTANT, masks_warped[i]);
}

vector<Mat> images_warped_f(imgs_.size());
for (size_t i = 0; i < imgs_.size(); ++i)
images_warped[i].convertTo(images_warped_f[i], CV_32F);

LOGLN("Warping images, time: " << ((getTickCount() - t) / getTickFrequency()) << " sec");

// Find seams
exposure_comp_->feed(corners, images_warped, masks_warped);
seam_finder_->find(images_warped_f, corners, masks_warped);

// Release unused memory
seam_est_imgs_.clear();
images_warped.clear();
images_warped_f.clear();
masks.clear();

LOGLN("Compositing...");
#if ENABLE_LOG
t = getTickCount();
#endif

Mat img_warped, img_warped_s;
Mat dilated_mask, seam_mask, mask, mask_warped;

//double compose_seam_aspect = 1;
double compose_work_aspect = 1;
bool is_blender_prepared = false;

double compose_scale = 1;
bool is_compose_scale_set = false;

Mat full_img, img;
for (size_t img_idx = 0; img_idx < imgs_.size(); ++img_idx)
{
LOGLN("Compositing image #" << indices_[img_idx] + 1);

// Read image and resize it if necessary
full_img = imgs_[img_idx];
if (!is_compose_scale_set)
{
if (compose_resol_ > 0)
compose_scale = min(1.0, sqrt(compose_resol_ * 1e6 / full_img.size().area()));
is_compose_scale_set = true;

// Compute relative scales
//compose_seam_aspect = compose_scale / seam_scale_;
compose_work_aspect = compose_scale / work_scale_;

// Update warped image scale
warped_image_scale_ *= static_cast<float>(compose_work_aspect);
w = warper_->create((float)warped_image_scale_);

// Update corners and sizes
for (size_t i = 0; i < imgs_.size(); ++i)
{
// Update intrinsics
cameras_[i].focal *= compose_work_aspect;
cameras_[i].ppx *= compose_work_aspect;
cameras_[i].ppy *= compose_work_aspect;

// Update corner and size
Size sz = full_img_sizes_[i];
if (std::abs(compose_scale - 1) > 1e-1)
{
sz.width = cvRound(full_img_sizes_[i].width * compose_scale);
sz.height = cvRound(full_img_sizes_[i].height * compose_scale);
}

Mat K;
cameras_[i].K().convertTo(K, CV_32F);
Rect roi = w->warpRoi(sz, K, cameras_[i].R);
corners[i] = roi.tl();
sizes[i] = roi.size();
}
}
if (std::abs(compose_scale - 1) > 1e-1)
resize(full_img, img, Size(), compose_scale, compose_scale);
else
img = full_img;
full_img.release();
Size img_size = img.size();

Mat K;
cameras_[img_idx].K().convertTo(K, CV_32F);

// Warp the current image
w->warp(img, K, cameras_[img_idx].R, INTER_LINEAR, BORDER_REFLECT, img_warped);

// Warp the current image mask
mask.create(img_size, CV_8U);
mask.setTo(Scalar::all(255));
w->warp(mask, K, cameras_[img_idx].R, INTER_NEAREST, BORDER_CONSTANT, mask_warped);

// Compensate exposure
exposure_comp_->apply((int)img_idx, corners[img_idx], img_warped, mask_warped);

img_warped.convertTo(img_warped_s, CV_16S);
img_warped.release();
img.release();
mask.release();

// Make sure seam mask has proper size
dilate(masks_warped[img_idx], dilated_mask, Mat());
resize(dilated_mask, seam_mask, mask_warped.size());

mask_warped = seam_mask & mask_warped;

if (!is_blender_prepared)
{
blender_->prepare(corners, sizes);
is_blender_prepared = true;
}

// Blend the current image
blender_->feed(img_warped_s, mask_warped, corners[img_idx]);
}

Mat result, result_mask;
blender_->blend(result, result_mask);

LOGLN("Compositing, time: " << ((getTickCount() - t) / getTickFrequency()) << " sec");

// Preliminary result is in CV_16SC3 format, but all values are in [0,255] range,
// so convert it to avoid user confusing
result.convertTo(pano_, CV_8U);

return OK;
}

stitching.cpp魚眼影象拼接融合 原始碼分析

相關文章

stitching.cpp魚眼影象拼接融合原始碼分析