【手撕算法】PatchMatch图像修复算法

/*********************************函数声明：初始化偏置图像参数：NONE注释：NONE测试：NONE**********************************/void PatchMatch::InitOff(Mat Mask, Mat &Off){  //为方便起见，将所有的都附上，要求不能赋值到非搜索区域  //初始化格式  Off = Mat(Mask.size(), CV_32FC2, Scalar::all(0));//2维无符号32位精度浮点数
  for (int i = 0; i < Mask.rows; i++)  {    for (int j = 0; j < Mask.cols; j++)    {      //不考虑search区域，没有破损，他们的最佳偏移向量当然是0，自己      if (Mask.at<uchar>(i, j) == search)      {        Off.at<Vec2f>(i, j)[0] = 0;  //<Vec2f> 向量，2维，浮点数        Off.at<Vec2f>(i, j)[1] = 0;      }      else//处理hole，采用随机偏置        {        //先初始化2个偏置数r_col,r_row        int r_col = rand() % Mask.cols; //rand（）产生随机数，主要是产生一个偏置的初始值        int r_row = rand() % Mask.rows;        r_col = r_col + j < Mask.cols ? r_col : r_col - Mask.cols;//边界检测        r_row = r_row + i < Mask.rows ? r_row : r_row - Mask.rows;
        //为什么要有这个循环？因为一次的随机赋值，很可能会出现偏置后的块跑到破损区域，或者是超出限定搜索框的边界        while (          !(Mask.at<uchar>(r_row + i, r_col + j) == search  //这里加上I，j，是因为他是A投影到B中的搜索偏置            && abs(r_row) < searchrowratio*Mask.rows))  //searchrowratio=0.5，搜索的时候,确保r_row偏置不会太远，一定是在原图像的大小里          {          r_col = rand() % Mask.cols;          r_row = rand() % Mask.rows;
          //边界检测          r_col = r_col + j < Mask.cols ? r_col : r_col - Mask.cols;          r_row = r_row + i < Mask.rows ? r_row : r_row - Mask.rows;        }
        //赋偏置值        Off.at<Vec2f>(i, j)[0] = r_row;        Off.at<Vec2f>(i, j)[1] = r_col;      }    }  }}

//以块为单位，用所有像素点的相同颜色通道的差平方来简单判断相似度float PatchMatch::Distance(Mat Dst, Mat Src){  float distance = 0;
  for (int i = 0; i < Dst.rows; i++)  {    for (int j = 0; j < Dst.cols; j++)    {      for (int k = 0; k < 3; k++)//K=3个颜色通道      {        int tem = Src.at < Vec3b >(i, j)[k] - Dst.at < Vec3b >(i, j)[k];        distance += tem * tem;//差平方      }    }  }
  return distance;}

//迭代第一步：传播//（now_row, now_col）：patch里的像素//odd：当前迭代次void PatchMatch::Propagation(Mat Dst, Mat Src, Mat Mask, Mat &Off, int row, int col,int odd){  Mat DstPatch = GetPatch(Dst, row, col);//获取长度为 patchsize = 3 的边界框, (row, col)代表的是中心像素点坐标
  if (odd % 2 == 0)//偶次迭代  {    //提取(row, col)的match块    Mat SrcPatch = GetPatch(Src, row + Off.at < Vec2f >(row, col)[0],      col + Off.at < Vec2f >(row, col)[1]);
    //提取(row, col-1)的match块    Mat LSrcPatch = GetPatch(Src, row + Off.at < Vec2f >(row, col - 1)[0],      col - 1 + Off.at < Vec2f >(row, col - 1)[1]);
    //提取(row-1, col)的match块    Mat USrcPatch = GetPatch(Src,      row - 1 + Off.at < Vec2f >(row - 1, col)[0],      col + Off.at < Vec2f >(row - 1, col)[1]);
    //返回上面4个块最相似的块的代表数字，用于switch判断    int location = GetMinPatch1(DstPatch, SrcPatch, LSrcPatch, USrcPatch);
    //利用上面的信息更新像素点的偏置地图    switch (location)    {      //若是1则不更新    case 2:      Off.at < Vec2f >(row, col)[0] = Off.at < Vec2f >(row, col - 1)[0];      Off.at < Vec2f >(row, col)[1] = Off.at < Vec2f >(row, col - 1)[1] - 1;      break;    case 3:      Off.at < Vec2f >(row, col)[0] = Off.at < Vec2f >(row - 1, col)[0] - 1;      Off.at < Vec2f >(row, col)[1] = Off.at < Vec2f >(row - 1, col)[1];      break;    }  }
  else//奇数次迭代  {    Mat SrcPatch = GetPatch(Src, row + Off.at < Vec2f >(row, col)[0],      col + Off.at < Vec2f >(row, col)[1]);    Mat RSrcPatch = GetPatch(Src, row + Off.at < Vec2f >(row, col + 1)[0],      col + 1 + Off.at < Vec2f >(row, col + 1)[1]);    Mat DSrcPatch = GetPatch(Src,      row + 1 + Off.at < Vec2f >(row + 1, col)[0],      col + Off.at < Vec2f >(row + 1, col)[1]);
    int location = GetMinPatch1(DstPatch, SrcPatch, RSrcPatch, DSrcPatch);    switch (location)    {    case 2:      Off.at < Vec2f >(row, col)[0] = Off.at < Vec2f >(row, col + 1)[0];      Off.at < Vec2f >(row, col)[1] = Off.at < Vec2f      >(row, col + 1)[1] + 1;      break;    case 3:      Off.at < Vec2f >(row, col)[0] = Off.at < Vec2f      >(row + 1, col)[0] + 1;      Off.at < Vec2f >(row, col)[1] = Off.at < Vec2f >(row + 1, col)[1];      break;    }  }}

//迭代第二步：随机搜索//（row,col）=（now_row, now_col）：修复patch里的像素void PatchMatch::RandomSearch(Mat Dst, Mat Src, Mat Mask, Mat &Off, int row, int col){  Mat DstPatch = GetPatch(Dst, row, col);//获取修复基准框，在框内操作
  //迭代指数  int attenuate = 0;
  while (true)  {    //获取随机参数，在 [-1;1] 间    float divcol = rand() % 2000 / 1000.0f - 1.0f;    float divrow = rand() % 2000 / 1000.0f - 1.0f;
    //减小框大小的公式，𝑢_𝑖=𝑣_0+𝑤*𝛼^𝑖*𝑅_𝑖    //行列分别处理，MaxWindow：原始框宽度；divcol：随机系数；pow(A,B):A的B次方。随迭代次数而变小的缩小系数；RandomAttenuation=0.5；    float veccol = MaxWindow * pow(RandomAttenuation, attenuate)* divcol;    float vecrow = MaxWindow * pow(RandomAttenuation, attenuate)* divrow;
    float length = sqrt(veccol * veccol + vecrow * vecrow);    //如果低于1个像素，没有意义，直接结束整个循环，对下一个像素处理    if (length < 1)      break;
    //x方向,前2项指向(row, col)的match块，后面是公式的后一项    int nowrow = row + Off.at < Vec2f >(row, col)[0] + vecrow;    //y方向    int nowcol = col + Off.at < Vec2f >(row, col)[1] + veccol;
    //判断随机搜索的patch不越界，在search内    if (nowcol >= 0 && nowcol <= Off.cols - 1 && nowrow >= 0      && nowrow <= Off.rows - 1      && Mask.at < uchar >(nowrow, nowcol) == search      && abs(nowrow - row) < searchrowratio * Mask.rows)//abs：绝对值    {      //取出原来的match块      Mat SrcPatch1 = GetPatch(Src, Off.at < Vec2f >(row, col)[0] + row,        Off.at < Vec2f >(row, col)[1] + col);      //取出现在的随机match块      Mat SrcPatch2 = GetPatch(Src, nowrow, nowcol);
      //对比相似性，找出最好的块      int location = GetMinPatch2(DstPatch, SrcPatch1, SrcPatch2);
      //结合最好的相似块给像素新的偏置值      switch (location)      {      case 2:        Off.at < Vec2f >(row, col)[1] = nowcol - col;        Off.at < Vec2f >(row, col)[0] = nowrow - row;        break;      }    }
    //迭代指数增加    attenuate++;  }}