GOICE项目初探

  1. 云栖社区>
  2. 博客>
  3. 正文

GOICE项目初探

禾路 2016-05-14 21:11:00 浏览366
展开阅读全文

       GOICE项目初探

        在图像拼接方面,市面上能够找到的软件中,要数MS的ICE效果、鲁棒性最好,而且界面也很美观。应该说有很多值得学习的地方,虽然这个项目不开源,但是利用现有的资料,也可以实现很多具体的拼接工作。
        

 

        基于现有的有限资源,主要是以opencv自己提供的stitch_detail进行修改和封包,基于ribbon编写界面,我也尝试实现了GOICE项目,实现全景图片的拼接、横向视频的拼接,如果下一步有时间的话再将双目实时拼接从以前的代码中移植过来。
        coding: https://coding.net/u/jsxyhelu/p/GOICE/git
这里简单地将一些技术要点进行解析,欢迎批评指正和合作交流!
一、对现有算法进行重新封装
   opencv原本的算法主要包含在Stitching Pipeline中,结构相对比较复杂,具体可以查看opencv refman
     对算法进行重构后整理如下:
   
 //使用变量
    Ptr<FeaturesFinder> finder;
    Mat full_img, img;
    int num_images = m_ImageList.size();
    vector<ImageFeatures> features(num_images);
    vector<Mat> images(num_images);
    vector<cv::Size> full_img_sizes(num_images);
    double seam_work_aspect = 1;
    vector<MatchesInfo> pairwise_matches;
    BestOf2NearestMatcher matcher(try_gpu, match_conf);
    vector<int> indices;
    vector<Mat> img_subset;
    vector<cv::Size> full_img_sizes_subset;
    HomographyBasedEstimator estimator;
    vector<CameraParams> cameras;
    vector<cv::Point> corners(num_images);
    vector<Mat> masks_warped(num_images);
    vector<Mat> images_warped(num_images);
    vector<cv::Size> sizes(num_images);
    vector<Mat> masks(num_images);
    Mat img_warped, img_warped_s;
    Mat dilated_mask, seam_mask, mask, mask_warped;
    Ptr<Blender> blender;
    double compose_work_aspect = 1;
 
    //拼接开始
    if (features_type == "surf")
        finder = new SurfFeaturesFinder();
    else
        finder = new OrbFeaturesFinder();
    
    //寻找特征点
    m_progress.SetPos(20);
    for (int i = 0; i < num_images; ++i)
    {
    
        full_img = m_ImageList[i].clone();
        full_img_sizes[i] = full_img.size();//读到的是大小
        if (full_img.empty())
        {
            MessageBox("图片读取错误,请确认后重新尝试!");
            return;
        }
        if (work_megapix < 0)
        {
            img = full_img;
            work_scale = 1;
            is_work_scale_set = true;
        }else{
            if (!is_work_scale_set)
            {
                work_scale = min(1.0, sqrt(work_megapix * 1e6 / full_img.size().area()));
                is_work_scale_set = true;
            }
            resize(full_img, img, cv::Size(), work_scale, work_scale);
        }
        if (!is_seam_scale_set)
        {
            seam_scale = min(1.0, sqrt(seam_megapix * 1e6 / full_img.size().area()));
            seam_work_aspect = seam_scale / work_scale;
            is_seam_scale_set = true;
        }
        (*finder)(img, features[i]);
        features[i].img_idx = i;
        resize(full_img, img, cv::Size(), seam_scale, seam_scale);
        images[i] = img.clone();
    }
 
    finder->collectGarbage();
    full_img.release();
    img.release();
    //进行匹配
    m_progress.SetPos(30);
    matcher(features, pairwise_matches);
    matcher.collectGarbage();
    indices  = leaveBiggestComponent(features, pairwise_matches, conf_thresh);
    for (size_t i = 0; i < indices.size(); ++i)
    {
        img_subset.push_back(images[indices[i]]);
        full_img_sizes_subset.push_back(full_img_sizes[indices[i]]);
    }
    m_progress.SetPos(40);
    images = img_subset;
 
    //判断图片是否足够
    num_images = static_cast<int>(img_subset.size());
    if (num_images < 2)
    {
        MessageBox("图片特征太少,尝试添加更多图片!");
        return;
    }
    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 intrinsics #" << indices[i]+1 << ":\n" << cameras[i].K());
    }
    //开始对准
    m_progress.SetPos(50);
    Ptr<detail::BundleAdjusterBase> adjuster;
    if (ba_cost_func == "reproj") adjuster = new detail::BundleAdjusterReproj();
    else 
    adjuster = new detail::BundleAdjusterRay();
    
    adjuster->setConfThresh(conf_thresh);
    Mat_<uchar> refine_mask = Mat::zeros(3, 3, CV_8U);
    if (ba_refine_mask[0] == 'x') refine_mask(0,0) = 1;
    if (ba_refine_mask[1] == 'x') refine_mask(0,1) = 1;
    if (ba_refine_mask[2] == 'x') refine_mask(0,2) = 1;
    if (ba_refine_mask[3] == 'x') refine_mask(1,1) = 1;
    if (ba_refine_mask[4] == 'x') refine_mask(1,2) = 1;
    adjuster->setRefinementMask(refine_mask);
    (*adjuster)(features, pairwise_matches, cameras);
 
    // Find median focal length
    vector<double> focals;
    for (size_t i = 0; i < cameras.size(); ++i)
        focals.push_back(cameras[i].focal);
    sort(focals.begin(), focals.end());
    float warped_image_scale;
    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;
    //开始融合
    m_progress.SetPos(60);
    if (do_wave_correct)
    {
        vector<Mat> rmats;
        for (size_t i = 0; i < cameras.size(); ++i)
            rmats.push_back(cameras[i].R);
        waveCorrect(rmats, wave_correct);
        for (size_t i = 0; i < cameras.size(); ++i)
            cameras[i].R = rmats[i];
    }
 
    //最后修正
    m_progress.SetPos(70);
    // Preapre images masks
    for (int i = 0; i < num_images; ++i)
    {
        masks[i].create(images[i].size(), CV_8U);
        masks[i].setTo(Scalar::all(255));
    }
    Ptr<WarperCreator> warper_creator;
    {
        if (warp_type == "plane") warper_creator = new cv::PlaneWarper();
        else if (warp_type == "cylindrical") warper_creator = new cv::CylindricalWarper();
        else if (warp_type == "spherical") warper_creator = new cv::SphericalWarper();
        else if (warp_type == "fisheye") warper_creator = new cv::FisheyeWarper();
        else if (warp_type == "stereographic") warper_creator = new cv::StereographicWarper();
        else if (warp_type == "compressedPlaneA2B1") warper_creator = new cv::CompressedRectilinearWarper(2, 1);
        else if (warp_type == "compressedPlaneA1.5B1") warper_creator = new cv::CompressedRectilinearWarper(1.5, 1);
        else if (warp_type == "compressedPlanePortraitA2B1") warper_creator = new cv::CompressedRectilinearPortraitWarper(2, 1);
        else if (warp_type == "compressedPlanePortraitA1.5B1") warper_creator = new cv::CompressedRectilinearPortraitWarper(1.5, 1);
        else if (warp_type == "paniniA2B1") warper_creator = new cv::PaniniWarper(2, 1);
        else if (warp_type == "paniniA1.5B1") warper_creator = new cv::PaniniWarper(1.5, 1);
        else if (warp_type == "paniniPortraitA2B1") warper_creator = new cv::PaniniPortraitWarper(2, 1);
        else if (warp_type == "paniniPortraitA1.5B1") warper_creator = new cv::PaniniPortraitWarper(1.5, 1);
        else if (warp_type == "mercator") warper_creator = new cv::MercatorWarper();
        else if (warp_type == "transverseMercator") warper_creator = new cv::TransverseMercatorWarper();
    }
    if (warper_creator.empty())
    {
        cout << "Can't create the following warper '" << warp_type << "'\n";
        return;}
 
    Ptr<RotationWarper> warper = warper_creator->create(static_cast<float>(warped_image_scale * seam_work_aspect));
 
    for (int i = 0; i < num_images; ++i)
    {
        Mat_<float> K;
        cameras[i].K().convertTo(K, CV_32F);
        float swa = (float)seam_work_aspect;
        K(0,0) *= swa; K(0,2) *= swa;
        K(1,1) *= swa; K(1,2) *= swa;
 
        corners[i] = warper->warp(images[i], K, cameras[i].R, INTER_LINEAR, BORDER_REFLECT, images_warped[i]);
        sizes[i] = images_warped[i].size();
 
        warper->warp(masks[i], K, cameras[i].R, INTER_NEAREST, BORDER_CONSTANT, masks_warped[i]);
    }
    vector<Mat> images_warped_f(num_images);
    for (int i = 0; i < num_images; ++i)
        images_warped[i].convertTo(images_warped_f[i], CV_32F);
    Ptr<ExposureCompensator> compensator = ExposureCompensator::createDefault(expos_comp_type);
    compensator->feed(corners, images_warped, masks_warped);
    //接缝修正
    m_progress.SetPos(80);
    Ptr<SeamFinder> seam_finder;
    if (seam_find_type == "no")
        seam_finder = new detail::NoSeamFinder();
    else if (seam_find_type == "voronoi")
        seam_finder = new detail::VoronoiSeamFinder();
    else if (seam_find_type == "gc_color")
        seam_finder = new detail::GraphCutSeamFinder(GraphCutSeamFinderBase::COST_COLOR);
    else if (seam_find_type == "gc_colorgrad")
        seam_finder = new detail::GraphCutSeamFinder(GraphCutSeamFinderBase::COST_COLOR_GRAD);
    else if (seam_find_type == "dp_color")
        seam_finder = new detail::DpSeamFinder(DpSeamFinder::COLOR);
    else if (seam_find_type == "dp_colorgrad")
        seam_finder = new detail::DpSeamFinder(DpSeamFinder::COLOR_GRAD);
    if (seam_finder.empty())
    {
        MessageBox("无法对图像进行缝隙融合");
        return;
    }
    //输出最后结果
    m_progress.SetPos(90);
    seam_finder->find(images_warped_f, corners, masks_warped);
    // Release unused memory
    images.clear();
    images_warped.clear();
    images_warped_f.clear();
    masks.clear();
 
    for (int img_idx = 0; img_idx < num_images; ++img_idx)
    {
        // Read image and resize it if necessary
        full_img = m_ImageList[img_idx];
        if (!is_compose_scale_set)
        {
            if (compose_megapix > 0)
                compose_scale = min(1.0, sqrt(compose_megapix * 1e6 / full_img.size().area()));
            is_compose_scale_set = true;
            // Compute relative scales
            compose_work_aspect = compose_scale / work_scale;
            // Update warped image scale
            warped_image_scale *= static_cast<float>(compose_work_aspect);
            warper = warper_creator->create(warped_image_scale);
            // Update corners and sizes
            for (int i = 0; i < num_images; ++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
                cv::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);
                cv::Rect roi = warper->warpRoi(sz, K, cameras[i].R);
                corners[i] = roi.tl();
                sizes[i] = roi.size();
            }
        }
        if (abs(compose_scale - 1) > 1e-1)
            resize(full_img, img, cv::Size(), compose_scale, compose_scale);
        else
            img = full_img;
        full_img.release();
        cv::Size img_size = img.size();
 
        Mat K;
        cameras[img_idx].K().convertTo(K, CV_32F);
 
        // Warp the current image
        warper->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));
        warper->warp(mask, K, cameras[img_idx].R, INTER_NEAREST, BORDER_CONSTANT, mask_warped);
 
        // Compensate exposure
        compensator->apply(img_idx, corners[img_idx], img_warped, mask_warped);
 
        img_warped.convertTo(img_warped_s, CV_16S);
        img_warped.release();
        img.release();
        mask.release();
 
        dilate(masks_warped[img_idx], dilated_mask, Mat());
        resize(dilated_mask, seam_mask, mask_warped.size());
        mask_warped = seam_mask & mask_warped;
 
        if (blender.empty())
        {
            blender = Blender::createDefault(blend_type, try_gpu);
            cv::Size dst_sz = resultRoi(corners, sizes).size();
            float blend_width = sqrt(static_cast<float>(dst_sz.area())) * blend_strength / 100.f;
            if (blend_width < 1.f)
                blender = Blender::createDefault(Blender::NO, try_gpu);
            else if (blend_type == Blender::MULTI_BAND)
            {
                MultiBandBlender* mb = dynamic_cast<MultiBandBlender*>(static_cast<Blender*>(blender));
                mb->setNumBands(static_cast<int>(ceil(log(blend_width)/log(2.)) - 1.));
                LOGLN("Multi-band blender, number of bands: " << mb->numBands());
            }
            else if (blend_type == Blender::FEATHER)
            {
                FeatherBlender* fb = dynamic_cast<FeatherBlender*>(static_cast<Blender*>(blender));
                fb->setSharpness(1.f/blend_width);
                LOGLN("Feather blender, sharpness: " << fb->sharpness());
            }
            blender->prepare(corners, sizes);
        }
 
        // Blend the current image
        blender->feed(img_warped_s, mask_warped, corners[img_idx]);
    }
    Mat result, result_mask;
    blender->blend(result, result_mask);
    m_progress.SetPos(100);
    AfxMessageBox("拼接成功!");
    m_progress.ShowWindow(false);
    m_progress.SetPos(0);
    //格式转换
    result.convertTo(result,CV_8UC3);
    showImage(result,IDC_PBDST);
    //保存结果
    m_matResult = result.clone();
      基本上没有修改代码的结构,但是做了几个改变
      1、原来的算法既读取文件名,又保存mat变量,我这里将其统一成为使用vector<Mat>来进行保存;
      2、将LOGLN的部分以messagebox的方式显示出来,并且进行错误控制;
      3、添加适当注释,并且在合适的地方控制进度条显示。
二、主要界面编写技巧
       主要界面使用了Ribbon的方法,结合使用IconWorkshop生成图标。如何生成这样的图片在我的博客中有专门介绍。
       内容方面,使用了基于listctrl的缩略图的显示,具体参考我的另一篇blog--"图像处理界面--缩略图的显示"
三、视频拼接的处理方法
       相比较图像拼接,这次添加了一个“横向视频”的拼接。其实算法原理是比较朴素的(当然这里考虑的是比较简单的情况)。就是对于精心拍摄的视频,那么只要每隔一段时间取一个图片,然后把这些图片进行拼接,就能够得到视频的全景图片。
void CMFCApplication1View::OnButtonOpenmov()
{
    CString pathName; 
    CString szFilters= _T("*(*.*)|*.*|avi(*.avi)|*.avi|mp4(*.mp4)|*.mp4||");
    CFileDialog dlg(TRUE,NULL,NULL,NULL,szFilters,this);
    VideoCapture capture;
    Mat frame;
    int iFrameCount = 0;
    int iFram = 0;
    if(dlg.DoModal()==IDOK){
            //获得路径
            pathName=dlg.GetPathName(); 
            //设置窗体
            m_ListThumbnail.ShowWindow(false);
            m_imagerect.ShowWindow(false);
            m_imagedst.ShowWindow(true);
            m_progress.ShowWindow(false);
            m_msg.ShowWindow(false);
            //打开视频并且抽取图片
            capture.open((string)pathName);
            if (!capture.isOpened())
            {
                MessageBox("视频打开错误!");
                return;
            }
            m_VectorMovImageNames.clear();
            m_MovImageList.clear();
            char cbuf[100];
            while (capture.read(frame))
            {
                //每隔50帧取一图
                if (0 == iFram%50)
                {
                    m_MovImageList.push_back(frame.clone());
                }
                showImage(frame,IDC_PBDST);
                iFram = iFram +1;
            }
    }
}
四、反思和小结
1)虽然现在已经对opencv的算法进行了集成,但是由于算法原理还是繁琐复杂的,下一步要结合对更复杂问题的进一步研究吃透算法;
2)使用ribbon进行程序设计现在已经比较熟悉了。能够认识到工具擅长解决的问题、能够认识到工具不好解决的问题,能够快速实现,才算是掌握;
 

 



目前方向:图像拼接融合、图像识别 联系方式:jsxyhelu@foxmail.com

网友评论

登录后评论
0/500
评论
禾路
+ 关注