1 #include <iostream>
2 #include <fstream>
3 #include <string>
4 #include "opencv2/opencv_modules.hpp"
5 #include <opencv2/core/utility.hpp>
6 #include "opencv2/imgcodecs.hpp"
7 #include "opencv2/highgui.hpp"
8 #include "opencv2/stitching/detail/autocalib.hpp"
9 #include "opencv2/stitching/detail/blenders.hpp"
10 #include "opencv2/stitching/detail/timelapsers.hpp"
11 #include "opencv2/stitching/detail/camera.hpp"
12 #include "opencv2/stitching/detail/exposure_compensate.hpp"
13 #include "opencv2/stitching/detail/matchers.hpp"
14 #include "opencv2/stitching/detail/motion_estimators.hpp"
15 #include "opencv2/stitching/detail/seam_finders.hpp"
16 #include "opencv2/stitching/detail/warpers.hpp"
17 #include "opencv2/stitching/warpers.hpp"
18 #ifdef HAVE_OPENCV_XFEATURES2D
19 #include "opencv2/xfeatures2d/nonfree.hpp"
20 #endif
21 #define ENABLE_LOG 1
22 #define LOG(msg) std::cout << msg
23 #define LOGLN(msg) std::cout << msg << std::endl
24 using namespace std;
25 using namespace cv;
26 using namespace cv::detail;
27 // Default command line args
28
29 #if 1
30 #define DLL_API __declspec(dllexport)
31 #else
32 #define DLL_API __declspec(dllimport)
33 #endif
34
35
36 extern "C" { //由于编译过程的原因，python一般只支持c的接口
37 typedef struct ImageBase {
38 int w; //图像的宽
39 int h; //图像的高
40 int c; //通道数
41 unsigned char *data; //我们要写python和c++交互的数据结构，0-255的单字符指针
42 }ImageMeta;
43 //typedef ImageBase ImageMeta;
44
45 DLL_API int Stitch(ImageMeta *im1, ImageMeta *im2);//函数导出，要改
46
47 };
48
49 //vector<String> img_names;
50 int num_images;
51 bool preview = false;
52 bool try_cuda = false;
53 double work_megapix = 0.6;
54 double seam_megapix = 0.1;
55 double compose_megapix = -1;
56 float conf_thresh = 1.f;
57 #ifdef HAVE_OPENCV_XFEATURES2D
58 string features_type = "surf";
59 #else
60 string features_type = "orb";
61 #endif
62 string matcher_type = "homography";
63 string estimator_type = "homography";
64 string ba_cost_func = "ray";
65 string ba_refine_mask = "xxxxx";
66 bool do_wave_correct = true;
67 WaveCorrectKind wave_correct = detail::WAVE_CORRECT_HORIZ;
68 bool save_graph = false;
69 std::string save_graph_to;
70 string warp_type = "spherical";
71 int expos_comp_type = ExposureCompensator::GAIN_BLOCKS;
72 int expos_comp_nr_feeds = 1;
73 int expos_comp_nr_filtering = 2;
74 int expos_comp_block_size = 32;
75 float match_conf = 0.3f;
76 string seam_find_type = "gc_color";
77 int blend_type = Blender::MULTI_BAND;
78 int timelapse_type = Timelapser::AS_IS;//延时摄影
79 float blend_strength = 5;
80 string result_name = "D:/result.jpg";//
81 bool timelapse = false;//首先定义timelapse的默认布尔类型为False
82 int range_width = -1;
83
84 DLL_API int Stitch(ImageMeta *im1, ImageMeta *im2)//入参两个数组指针，出参一个数组指针
85 //vector<Mat> img_list
86 {//一个int数；一个图片类型的列表
87 //predict先判断长度 然后长度作为一个参数传给
88 //preview = true;
89 //try_cuda = true;
90 //preview = true;
91 //result = 'D:/result.jpg';
92 //work_megapix = -1;
93 //features_type = "orb";
94
95 Mat img1 = Mat::zeros(Size(im1->w, im1->h), CV_8UC3);
96 //先从输入的指针对象提取w,h，data；将python传来的参数转变成C处理的格式。用的是相同的结构：结构体。
97 img1.data = im1->data;
98
99 Mat img2 = Mat::zeros(Size(im2->w, im2->h), CV_8UC3);
100 //先从输入的指针对象提取w,h，data；将python传来的参数转变成C处理的格式。用的是相同的结构：结构体。
101 img2.data = im2->data;
102
103
104 //Mat img1, img2;
105 //img1 = imread("D:/1Hill.jpg");
106 //img2 = imread("D:/2Hill.jpg");
107 vector<Mat> ALLimages(2);
108 ALLimages[0] = img1.clone();
109 ALLimages[1] = img2.clone();
110 //img_names.push_back("D:/1Hill.jpg");
111 //img_names.push_back("D:/2Hill.jpg");//??
112 //img_names.push_back("D:/3Hill.jpg");//??
113 num_images = 2;
114 #if ENABLE_LOG
115 int64 app_start_time = getTickCount();
116 #endif
117 #if 0
118 cv::setBreakOnError(true);
119 #endif
120 //int retval = parseCmdArgs(argc, argv);
121 //if (retval)
122 //return retval;
123 // Check if have enough images
124 //int num_images = static_cast<int>(img_names.size());
125 if (num_images < 2)
126 {
127 LOGLN("Need more images");
128 return -1;
129 }
130 double work_scale = 1, seam_scale = 1, compose_scale = 1;
131 bool is_work_scale_set = false, is_seam_scale_set = false, is_compose_scale_set = false;
132 LOGLN("Finding features...");
133 #if ENABLE_LOG
134 int64 t = getTickCount();
135 #endif
136 Ptr<Feature2D> finder;
137 if (features_type == "orb")
138 {
139 finder = ORB::create();
140 }
141 else if (features_type == "akaze")
142 {
143 finder = AKAZE::create();
144 }
145 #ifdef HAVE_OPENCV_XFEATURES2D
146 else if (features_type == "surf")
147 {
148 finder = xfeatures2d::SURF::create();
149 }
150 else if (features_type == "sift") {
151 finder = xfeatures2d::SIFT::create();
152 }
153 #endif
154 else
155 {
156 cout << "Unknown 2D features type: '" << features_type << "'.\n";
157 return -1;
158 }
159 Mat full_img, img;
160 vector<ImageFeatures> features(num_images);
161 vector<Mat> images(num_images);
162 vector<Size> full_img_sizes(num_images);
163 double seam_work_aspect = 1;
164 for (int i = 0; i < num_images; ++i)
165 {
166 full_img = ALLimages[i];
167
168 //获取一张图。imread_img -------->Mat
169 //先把一边调通了再去组合调试，分治
170 //full_img = img_list;
171 //python传进来n张图片的base64,可以转成读取后的图片。
172
173 //先在c中定义图像HWC结构数组数组转一次 Mat, dll返回Mat结果，Mat转一次结构体
174 //main输入 Mat1,Mat2
175 //dll返回数组，python转化成cv2image,然后输出image2base64
176
177 //full_image里面是读取的imread_img类型
178 //base64的size容易确定
179 //先在predict前提取到图片的整个Mat传给DLL
180 full_img_sizes[i] = full_img.size();//结果：full_img_sizes = [(500,300),(200,100)]
181 if (full_img.empty())
182 {
183 //LOGLN("Can't open image " << img_names[i]);//访问了空指针，和img_names有关
184 return -2;
185 }
186 if (work_megapix < 0)
187 {
188 img = full_img;
189 work_scale = 1;
190 is_work_scale_set = true;
191 }
192 else
193 {
194 if (!is_work_scale_set)
195 {
196 work_scale = min(1.0, sqrt(work_megapix * 1e6 / full_img.size().area()));
197 is_work_scale_set = true;
198 }
199 resize(full_img, img, Size(), work_scale, work_scale, INTER_LINEAR_EXACT);
200 }
201 if (!is_seam_scale_set)
202 {
203 seam_scale = min(1.0, sqrt(seam_megapix * 1e6 / full_img.size().area()));
204 seam_work_aspect = seam_scale / work_scale;
205 is_seam_scale_set = true;
206 }
207 computeImageFeatures(finder, img, features[i]);
208 features[i].img_idx = i;
209 LOGLN("Features in image #" << i + 1 << ": " << features[i].keypoints.size());
210 resize(full_img, img, Size(), seam_scale, seam_scale, INTER_LINEAR_EXACT);
211 images[i] = img.clone();
212 //循环是为了找到每张图的特征，然后把图片copy到images里
213 }
214 full_img.release();
215 img.release();
216 LOGLN("Finding features, time: " << ((getTickCount() - t) / getTickFrequency()) << " sec");
217 LOG("Pairwise matching");
218 #if ENABLE_LOG
219 t = getTickCount();
220 #endif
221 vector<MatchesInfo> pairwise_matches;
222 Ptr<FeaturesMatcher> matcher;
223 if (matcher_type == "affine")
224 matcher = makePtr<AffineBestOf2NearestMatcher>(false, try_cuda, match_conf);
225 else if (range_width == -1)
226 matcher = makePtr<BestOf2NearestMatcher>(try_cuda, match_conf);
227 else
228 matcher = makePtr<BestOf2NearestRangeMatcher>(range_width, try_cuda, match_conf);
229 (*matcher)(features, pairwise_matches);
230 matcher->collectGarbage();
231 LOGLN("Pairwise matching, time: " << ((getTickCount() - t) / getTickFrequency()) << " sec");
232 // Check if we should save matches graph
233 //if (save_graph)
234 //{
235 // LOGLN("Saving matches graph...");
236 // ofstream f(save_graph_to.c_str());
237 // f << matchesGraphAsString(img_names, pairwise_matches, conf_thresh);
238 //}
239 // Leave only images we are sure are from the same panorama
240 vector<int> indices = leaveBiggestComponent(features, pairwise_matches, conf_thresh);
241 if (indices.size() != 2)//判断两个图片的相关性
242 return -1;
243
244 if (num_images < 2)
245 {
246 LOGLN("Need more images");
247 return -1;
248 }
249 Ptr<Estimator> estimator;
250 if (estimator_type == "affine")
251 estimator = makePtr<AffineBasedEstimator>();
252 else
253 estimator = makePtr<HomographyBasedEstimator>();
254 vector<CameraParams> cameras;
255 if (!(*estimator)(features, pairwise_matches, cameras))
256 {
257 cout << "Homography estimation failed.\n";
258 return -1;
259 }
260 for (size_t i = 0; i < cameras.size(); ++i)
261 {
262 Mat R;
263 cameras[i].R.convertTo(R, CV_32F);
264 cameras[i].R = R;
265 //LOGLN("Initial camera intrinsics #" << indices[i] + 1 << ":\nK:\n" << cameras[i].K() << "\nR:\n" << cameras[i].R);
266 }
267 Ptr<detail::BundleAdjusterBase> adjuster;
268 if (ba_cost_func == "reproj") adjuster = makePtr<detail::BundleAdjusterReproj>();
269 else if (ba_cost_func == "ray") adjuster = makePtr<detail::BundleAdjusterRay>();
270 else if (ba_cost_func == "affine") adjuster = makePtr<detail::BundleAdjusterAffinePartial>();
271 else if (ba_cost_func == "no") adjuster = makePtr<NoBundleAdjuster>();
272 else
273 {
274 cout << "Unknown bundle adjustment cost function: '" << ba_cost_func << "'.\n";
275 return -1;
276 }
277 adjuster->setConfThresh(conf_thresh);
278 Mat_<uchar> refine_mask = Mat::zeros(3, 3, CV_8U);
279 if (ba_refine_mask[0] == 'x') refine_mask(0, 0) = 1;
280 if (ba_refine_mask[1] == 'x') refine_mask(0, 1) = 1;
281 if (ba_refine_mask[2] == 'x') refine_mask(0, 2) = 1;
282 if (ba_refine_mask[3] == 'x') refine_mask(1, 1) = 1;
283 if (ba_refine_mask[4] == 'x') refine_mask(1, 2) = 1;
284 adjuster->setRefinementMask(refine_mask);
285 if (!(*adjuster)(features, pairwise_matches, cameras))
286 {
287 cout << "Camera parameters adjusting failed.\n";
288 return -1;
289 }
290 // Find median focal length
291 vector<double> focals;
292 for (size_t i = 0; i < cameras.size(); ++i)
293 {
294 //LOGLN("Camera #" << indices[i] + 1 << ":\nK:\n" << cameras[i].K() << "\nR:\n" << cameras[i].R);
295 focals.push_back(cameras[i].focal);
296 }
297 sort(focals.begin(), focals.end());
298 float warped_image_scale;
299 if (focals.size() % 2 == 1)
300 warped_image_scale = static_cast<float>(focals[focals.size() / 2]);
301 else
302 warped_image_scale = static_cast<float>(focals[focals.size() / 2 - 1] + focals[focals.size() / 2]) * 0.5f;
303 if (do_wave_correct)
304 {
305 vector<Mat> rmats;
306 for (size_t i = 0; i < cameras.size(); ++i)
307 rmats.push_back(cameras[i].R.clone());
308 waveCorrect(rmats, wave_correct);
309 for (size_t i = 0; i < cameras.size(); ++i)
310 cameras[i].R = rmats[i];
311 }
312 LOGLN("Warping images (auxiliary)... ");
313 #if ENABLE_LOG
314 t = getTickCount();
315 #endif
316 vector<Point> corners(num_images);
317 vector<UMat> masks_warped(num_images);
318 vector<UMat> images_warped(num_images);
319 vector<Size> sizes(num_images);
320 vector<UMat> masks(num_images);
321 // Prepare images masks
322 for (int i = 0; i < num_images; ++i)
323 {
324 masks[i].create(images[i].size(), CV_8U);
325 masks[i].setTo(Scalar::all(255));
326 }
327 // Warp images and their masks
328 Ptr<WarperCreator> warper_creator;
329 #ifdef HAVE_OPENCV_CUDAWARPING
330 if (try_cuda && cuda::getCudaEnabledDeviceCount() > 0)
331 {
332 if (warp_type == "plane")
333 warper_creator = makePtr<cv::PlaneWarperGpu>();
334 else if (warp_type == "cylindrical")
335 warper_creator = makePtr<cv::CylindricalWarperGpu>();
336 else if (warp_type == "spherical")
337 warper_creator = makePtr<cv::SphericalWarperGpu>();
338 }
339 else
340 #endif
341 {
342 if (warp_type == "plane")
343 warper_creator = makePtr<cv::PlaneWarper>();
344 else if (warp_type == "affine")
345 warper_creator = makePtr<cv::AffineWarper>();
346 else if (warp_type == "cylindrical")
347 warper_creator = makePtr<cv::CylindricalWarper>();
348 else if (warp_type == "spherical")
349 warper_creator = makePtr<cv::SphericalWarper>();
350 else if (warp_type == "fisheye")
351 warper_creator = makePtr<cv::FisheyeWarper>();
352 else if (warp_type == "stereographic")
353 warper_creator = makePtr<cv::StereographicWarper>();
354 else if (warp_type == "compressedPlaneA2B1")
355 warper_creator = makePtr<cv::CompressedRectilinearWarper>(2.0f, 1.0f);
356 else if (warp_type == "compressedPlaneA1.5B1")
357 warper_creator = makePtr<cv::CompressedRectilinearWarper>(1.5f, 1.0f);
358 else if (warp_type == "compressedPlanePortraitA2B1")
359 warper_creator = makePtr<cv::CompressedRectilinearPortraitWarper>(2.0f, 1.0f);
360 else if (warp_type == "compressedPlanePortraitA1.5B1")
361 warper_creator = makePtr<cv::CompressedRectilinearPortraitWarper>(1.5f, 1.0f);
362 else if (warp_type == "paniniA2B1")
363 warper_creator = makePtr<cv::PaniniWarper>(2.0f, 1.0f);
364 else if (warp_type == "paniniA1.5B1")
365 warper_creator = makePtr<cv::PaniniWarper>(1.5f, 1.0f);
366 else if (warp_type == "paniniPortraitA2B1")
367 warper_creator = makePtr<cv::PaniniPortraitWarper>(2.0f, 1.0f);
368 else if (warp_type == "paniniPortraitA1.5B1")
369 warper_creator = makePtr<cv::PaniniPortraitWarper>(1.5f, 1.0f);
370 else if (warp_type == "mercator")
371 warper_creator = makePtr<cv::MercatorWarper>();
372 else if (warp_type == "transverseMercator")
373 warper_creator = makePtr<cv::TransverseMercatorWarper>();
374 }
375 if (!warper_creator)
376 {
377 cout << "Can't create the following warper '" << warp_type << "'\n";
378 return 1;
379 }
380 Ptr<RotationWarper> warper = warper_creator->create(static_cast<float>(warped_image_scale * seam_work_aspect));
381 for (int i = 0; i < num_images; ++i)
382 {
383 Mat_<float> K;
384 cameras[i].K().convertTo(K, CV_32F);
385 float swa = (float)seam_work_aspect;
386 K(0, 0) *= swa; K(0, 2) *= swa;
387 K(1, 1) *= swa; K(1, 2) *= swa;
388 corners[i] = warper->warp(images[i], K, cameras[i].R, INTER_LINEAR, BORDER_REFLECT, images_warped[i]);
389 sizes[i] = images_warped[i].size();
390 warper->warp(masks[i], K, cameras[i].R, INTER_NEAREST, BORDER_CONSTANT, masks_warped[i]);
391 }
392 vector<UMat> images_warped_f(num_images);
393 for (int i = 0; i < num_images; ++i)
394 images_warped[i].convertTo(images_warped_f[i], CV_32F);
395 LOGLN("Warping images, time: " << ((getTickCount() - t) / getTickFrequency()) << " sec");
396 LOGLN("Compensating exposure...");
397 #if ENABLE_LOG
398 t = getTickCount();
399 #endif
400 Ptr<ExposureCompensator> compensator = ExposureCompensator::createDefault(expos_comp_type);
401 if (dynamic_cast<GainCompensator*>(compensator.get()))
402 {
403 GainCompensator* gcompensator = dynamic_cast<GainCompensator*>(compensator.get());
404 gcompensator->setNrFeeds(expos_comp_nr_feeds);
405 }
406 if (dynamic_cast<ChannelsCompensator*>(compensator.get()))
407 {
408 ChannelsCompensator* ccompensator = dynamic_cast<ChannelsCompensator*>(compensator.get());
409 ccompensator->setNrFeeds(expos_comp_nr_feeds);
410 }
411 if (dynamic_cast<BlocksCompensator*>(compensator.get()))
412 {
413 BlocksCompensator* bcompensator = dynamic_cast<BlocksCompensator*>(compensator.get());
414 bcompensator->setNrFeeds(expos_comp_nr_feeds);
415 bcompensator->setNrGainsFilteringIterations(expos_comp_nr_filtering);
416 bcompensator->setBlockSize(expos_comp_block_size, expos_comp_block_size);
417 }
418 compensator->feed(corners, images_warped, masks_warped);
419 LOGLN("Compensating exposure, time: " << ((getTickCount() - t) / getTickFrequency()) << " sec");
420 LOGLN("Finding seams...");
421 #if ENABLE_LOG
422 t = getTickCount();
423 #endif
424 Ptr<SeamFinder> seam_finder;
425 if (seam_find_type == "no")
426 seam_finder = makePtr<detail::NoSeamFinder>();
427 else if (seam_find_type == "voronoi")
428 seam_finder = makePtr<detail::VoronoiSeamFinder>();
429 else if (seam_find_type == "gc_color")
430 {
431 #ifdef HAVE_OPENCV_CUDALEGACY
432 if (try_cuda && cuda::getCudaEnabledDeviceCount() > 0)
433 seam_finder = makePtr<detail::GraphCutSeamFinderGpu>(GraphCutSeamFinderBase::COST_COLOR);
434 else
435 #endif
436 seam_finder = makePtr<detail::GraphCutSeamFinder>(GraphCutSeamFinderBase::COST_COLOR);
437 }
438 else if (seam_find_type == "gc_colorgrad")
439 {
440 #ifdef HAVE_OPENCV_CUDALEGACY
441 if (try_cuda && cuda::getCudaEnabledDeviceCount() > 0)
442 seam_finder = makePtr<detail::GraphCutSeamFinderGpu>(GraphCutSeamFinderBase::COST_COLOR_GRAD);
443 else
444 #endif
445 seam_finder = makePtr<detail::GraphCutSeamFinder>(GraphCutSeamFinderBase::COST_COLOR_GRAD);
446 }
447 else if (seam_find_type == "dp_color")
448 seam_finder = makePtr<detail::DpSeamFinder>(DpSeamFinder::COLOR);
449 else if (seam_find_type == "dp_colorgrad")
450 seam_finder = makePtr<detail::DpSeamFinder>(DpSeamFinder::COLOR_GRAD);
451 if (!seam_finder)
452 {
453 cout << "Can't create the following seam finder '" << seam_find_type << "'\n";
454 return 1;
455 }
456 seam_finder->find(images_warped_f, corners, masks_warped);
457 LOGLN("Finding seams, time: " << ((getTickCount() - t) / getTickFrequency()) << " sec");
458 // Release unused memory
459 images.clear();
460 images_warped.clear();
461 images_warped_f.clear();
462 masks.clear();
463 LOGLN("Compositing...");
464 #if ENABLE_LOG
465 t = getTickCount();
466 #endif
467 Mat img_warped, img_warped_s;
468 Mat dilated_mask, seam_mask, mask, mask_warped;
469 Ptr<Blender> blender;
470 Ptr<Timelapser> timelapser;
471 //double compose_seam_aspect = 1;
472 double compose_work_aspect = 1;
473 for (int img_idx = 0; img_idx < num_images; ++img_idx)
474 {
475 //LOGLN("Compositing image #" << indices[img_idx] + 1);
476 // Read image and resize it if necessary
477 full_img = ALLimages[img_idx];
478 if (!is_compose_scale_set)
479 {
480 if (compose_megapix > 0)
481 compose_scale = min(1.0, sqrt(compose_megapix * 1e6 / full_img.size().area()));
482 is_compose_scale_set = true;
483 // Compute relative scales
484 //compose_seam_aspect = compose_scale / seam_scale;
485 compose_work_aspect = compose_scale / work_scale;
486 // Update warped image scale
487 warped_image_scale *= static_cast<float>(compose_work_aspect);
488 warper = warper_creator->create(warped_image_scale);
489 // Update corners and sizes
490 for (int i = 0; i < num_images; ++i)
491 {
492 // Update intrinsics
493 cameras[i].focal *= compose_work_aspect;
494 cameras[i].ppx *= compose_work_aspect;
495 cameras[i].ppy *= compose_work_aspect;
496 // Update corner and size
497 Size sz = full_img_sizes[i];
498 Mat K;
499 cameras[i].K().convertTo(K, CV_32F);
500 Rect roi = warper->warpRoi(sz, K, cameras[i].R);
501 corners[i] = roi.tl();
502 sizes[i] = roi.size();
503 }
504 }
505 if (abs(compose_scale - 1) > 1e-1)//没用
506 resize(full_img, img, Size(), compose_scale, compose_scale, INTER_LINEAR_EXACT);
507 else
508 img = full_img;
509 full_img.release();
510 Size img_size = img.size();
511 Mat K;
512 cameras[img_idx].K().convertTo(K, CV_32F);
513 // Warp the current image
514 warper->warp(img, K, cameras[img_idx].R, INTER_LINEAR, BORDER_REFLECT, img_warped);
515 // Warp the current image mask
516 mask.create(img_size, CV_8U);
517 mask.setTo(Scalar::all(255));
518 warper->warp(mask, K, cameras[img_idx].R, INTER_NEAREST, BORDER_CONSTANT, mask_warped);
519 // Compensate exposure
520 compensator->apply(img_idx, corners[img_idx], img_warped, mask_warped);
521 img_warped.convertTo(img_warped_s, CV_16S);
522 img_warped.release();
523 img.release();
524 mask.release();
525 dilate(masks_warped[img_idx], dilated_mask, Mat());
526 resize(dilated_mask, seam_mask, mask_warped.size(), 0, 0, INTER_LINEAR_EXACT);
527 mask_warped = seam_mask & mask_warped;
528 if (!blender && !timelapse)//blender是False,timelapse也是False,这里运行了！
529 {//做multiband
530 blender = Blender::createDefault(blend_type, try_cuda);
531 Size dst_sz = resultRoi(corners, sizes).size();
532 float blend_width = sqrt(static_cast<float>(dst_sz.area())) * blend_strength / 100.f;
533 if (blend_width < 1.f)
534 blender = Blender::createDefault(Blender::NO, try_cuda);
535 else if (blend_type == Blender::MULTI_BAND)
536 {
537 MultiBandBlender* mb = dynamic_cast<MultiBandBlender*>(blender.get());
538 mb->setNumBands(static_cast<int>(ceil(log(blend_width) / log(2.)) - 1.));
539 LOGLN("Multi-band blender, number of bands: " << mb->numBands());
540 }
541 else if (blend_type == Blender::FEATHER)//未运行
542 {
543 FeatherBlender* fb = dynamic_cast<FeatherBlender*>(blender.get());
544 fb->setSharpness(1.f / blend_width);
545 LOGLN("Feather blender, sharpness: " << fb->sharpness());
546 }
547 blender->prepare(corners, sizes);
548 }
549 else if (!timelapser && timelapse)//timelapse是假，timelapser是什么？？没运行
550 {
551 timelapser = Timelapser::createDefault(timelapse_type);
552 timelapser->initialize(corners, sizes);
553 cout << "----------------------------运行---------------------------------" << endl;
554 }
555 // Blend the current image
556 if (timelapse)//默认是假
557 {
558 cout << "----------------------------运行2---------------------------------" << endl;
559 }
560 else
561 {//这里运行了两次，因为在循环体中，图片有两张
562 blender->feed(img_warped_s, mask_warped, corners[img_idx]);
563 cout << "----------------------------运行3---------------------------------" << endl;
564 }
565 }
566 if (!timelapse)//运行了
567 {
568 Mat result, result_mask;
569 blender->blend(result, result_mask);
570 LOGLN("Compositing, time: " << ((getTickCount() - t) / getTickFrequency()) << " sec");
571 imwrite(result_name, result);
572 //(result.cols)*(result.rows)
573 memcpy(im2->data, result.clone().data, (result.cols)*(result.rows));//从哪里拷贝多少个字节。。
574 im2->w = result.cols;
575 im2->h = result.rows;
576 im2->c = 3;
577 }
578 LOGLN("Finished, total time: " << ((getTickCount() - app_start_time) / getTickFrequency()) << " sec");
579 return 0;
580 }
1 from ctypes import *
2 from io import BytesIO
3 import numpy as np
4 import cv2
5
6
7 # 写法是yolo的darknet.py里的
8
9 def c_array(ctype, values): # 把图像的数据转化为内存连续的 列表 ， 使c++能使用这块内存
10 arr = (ctype * len(values))()
11 arr[:] = values
12 return arr
13
14
15 def array_to_image(arr):
16 c = arr.shape[2]
17 h = arr.shape[0]
18 w = arr.shape[1]
19 arr = arr.flatten()#转化成图片后成了一维的
20 data = c_array(c_uint8, arr)
21 im = IMAGE(w, h, c, data)#将读进来数组转化成c接受的形式，调用class IMAGE
22 return im
23
24
25 class IMAGE(Structure): # 这里和ImgSegmentation.hpp里面的结构体保持一致。
26 _fields_ = [("w", c_int),
27 ("h", c_int),
28 ("c", c_int),
29 ("data", POINTER(c_uint8))]
30
31
32 img1 = cv2.imread('D:/1Hill.jpg')
33 img2 = cv2.imread('D:/2Hill.jpg')
34 #h, w, c = img.shape[0], img.shape[1], img.shape[2]
35 #h, w, c = img.shape[0], img.shape[1], img.shape[2]
36
37 #gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY) * 0
38 #gray = np.reshape(gray, (h, w, 1)) # 一定要使用(h, w, 1)，最后的1别忘。
39 im1 = array_to_image(img1)#这里是将读进来的cv_imread格式图片转化成结构体，一维的
40 im2 = array_to_image(img2)
41 #gray_img = array_to_image(gray)
42
43 lib = cdll.LoadLibrary('./image_stiching.dll') # 读取动态库文件
44 lib.Stitch.argtypes = [POINTER(IMAGE), POINTER(IMAGE)] # 设置函数入参格式,声明采用指针传递。指定入参为2个数组指针，python里定义类型指针和C相反，类型在后。
45 #lib.Stitch.restype = c_int64
46 lib.Stitch(im1, im2) # 执行函数，这里直接修改gray_img的内存数据。入参是非指针，python提取地址作为输入。因为函数原型是传递的指针，这里相当于POINTER会自动取输入im1,im2的指针作为入参。
47 ## 因此输入的内存数据会直接被改变。
48 y = im2.data # 获取data，被改变传递改变的对象名
49 array_length = im2.h * im2.w
50 #转化为numpy的ndarray
51 buffer_from_memory = pythonapi.PyMemoryView_FromMemory # 这个是python 3的使用方法，提取运算缓存
52 buffer_from_memory.restype = py_object #提取缓存返回的数据格式，以上两步是下一步从缓存中提取某个变量的结果必须的。
53 buffer = buffer_from_memory(y, array_length) #提取底层的缓存指针，指定提取缓存大小
54 img = np.frombuffer(buffer, dtype=np.uint8) #提取到缓存中的数组
55 print("----------------------")
56 print(img.shape)
57 img = np.reshape(img, (im2.h, im2.w,1)) #改变缓存数组的格式，用于显示
58 print("-------2---------")
59 print(img.shape)
60 print(img)
61 cv2.imshow('test', img)
62 cv2.imwrite("D:/RESULT_PY.JPG",img)
63 cv2.waitKey(0)
64
