UMHexagonS搜索过程

通过相邻块的预测得到mvp后，会以mvp为基础搜索最佳的匹配块，UMHexagonS就是h.264中用的一种搜索算法。

UMHexagonS是一种整像素搜索算法，也就是搜索过程中，参考图像一直都是原来的重构图像，并没有使用经过插值的图像进行搜索。

首先UMHexagonS会根据相关信息去得到比较有可能的mv，（然后用小菱形搜索到该区域去搜索该区域中的最佳mv，这种情况会在下面注明，至于如何才算最佳，请参照http://www.cnblogs.com/TaigaCon/p/3790218.html）

由于UMHexagonS是一种整像素搜索算法，所以会存在对分数的mv取整的情况，此时取整是指把mv对齐到某个像素上，消去分数部分

首先需要选取合适的搜索起点，有以下几种起点的选择

1.mvp

由于还是整像素搜索，所以这里需要对mvp取整，得到的整数的mv后采用小菱形搜索以得到比较优秀的mv。

2.原点

原点，即mv为0，即当前块的位置（然后采用小菱形搜索）

3.上层块mv

参考下图，如果当前块为8x8，那么覆盖当前块的16x8块就是其上层块

运动搜索中，分块模式有7种

模式4的上层模式为2，模式7的上层模式为4

4.共同位置块mv，取上一参考图像与当前块相同位置的块的mv,然后取整

5.共同位置参考mv通过参考图像距离计算后得到的mv，然后取整

6.最后还采用一次小菱形搜索

这里的小菱形搜索主要为了对上面3、4、5预测后得到的最佳mv再采用一次小菱形搜索以得到该区域内最佳mv

小菱形搜索就是把mv的x,y分别+1，-1后得到的新mv，然后各自对比得到其中最优的mv

Early Termination

通过上面的步骤得到最优的搜索起点后，需要计算该mv的匹配满意程度，以跳转做不同的后续搜索处理，该过程叫Early Termination。

Early Termination由于涉及到数学上的分析，所以会在后面的章节再细述。

Early Termination有两个个跳转出口，分别代表不同的匹配满意程度：

• Extended Hexagon-based Search（六边形模板反复搜索）               满意
• the third step with a small search pattern（小菱形模板反复搜索）   很满意

但是如果在不甚满意的情况下，Early Termination会不作跳转，直接执行下一步

UMH搜索

经过上面步骤后，得到其中最佳的搜索起点的mv，如果该mv经Early Termination判断为不甚满意，会以该mv为中心，直接开始UMH搜索。UMH搜索有以下步骤

1.Unsymmetrical-cross search（非对称十字搜索）

非对称十字搜索会先后对x轴与y轴进行搜索，y轴的搜索范围是x轴的一半，这是因为在一般的视频中，镜头的纵向移动距离会比较短，横向移动距离会比较长，而且比较常见。搜索时，横轴的搜索范围是search range，而纵轴会是它的一般。

2.Spiral search（螺旋搜索）

螺旋搜索采用的是full search（全搜索）的搜索方法，但是搜索步长只有24，相当于5x5的区域。而全搜索会对整个搜索范围进行搜索。

3.Uneven Multi-Hexagon-grid Search（不规律六边形模板搜索）

这种搜索方式是以当前mv指向的像素点为圆心，一圈一圈地往外搜索，一旦在某个圈内搜索到更佳的位置，立刻停止搜索，否则搜索完整个搜索范围

4.Extended Hexagon-based Search（六边形模板反复搜索）

不同于上一个搜索方式，这种搜索方式是以当前最佳mv指向的像素点为圆心，进行一次六边形模板搜索，一旦搜索到某个更佳的位置，则以此位置为圆心，重新进行一次六边形模板搜索。如果没有比圆心更佳的位置，则终止搜索。

5.the third step with a small search pattern（小菱形模板反复搜索）

类似Extended Hexagon-based Search（六边形模板反复搜索）的搜索方式，不过把六边形换成了菱形

以上可参照jvt-G016

JM8.6

/*!
 ************************************************************************
 * \brief用非对称十字形多层次六边形格点搜索算法进行运动搜索
 *    FastIntegerPelBlockMotionSearch: fast pixel block motion search
 *    this algrithm is called UMHexagonS(see JVT-D016),which includes
 *    four steps with different kinds of search patterns
 * \par Input:
 * pel_t**   orig_pic,     // <--  original picture
 * int       ref,          // <--  reference frame (0... or -1 (backward))
 * int       pic_pix_x,    // <--  absolute x-coordinate of regarded AxB block
 * int       pic_pix_y,    // <--  absolute y-coordinate of regarded AxB block
 * int       blocktype,    // <--  block type (1-16x16 ... 7-4x4)
 * int       pred_mv_x,    // <--  motion vector predictor (x) in sub-pel units
 * int       pred_mv_y,    // <--  motion vector predictor (y) in sub-pel units
 * int*      mv_x,         //  --> motion vector (x) - in pel units
 * int*      mv_y,         //  --> motion vector (y) - in pel units
 * int       search_range, // <--  1-d search range in pel units
 * int       min_mcost,    // <--  minimum motion cost (cost for center or huge value)
 * double    lambda        // <--  lagrangian parameter for determining motion cost
 * \par
 * Three macro definitions defined in this program:
 * 1. EARLY_TERMINATION: early termination algrithm, refer to JVT-D016.doc
 * 2. SEARCH_ONE_PIXEL: search one pixel in search range
 * 3. SEARCH_ONE_PIXEL1(value_iAbort): search one pixel in search range,
 *                                 but give a parameter to show if mincost refeshed
 * \ Main contributors: (see contributors.h for copyright, address and affiliation details)
 *   Zhibo Chen         <chenzhibo@tsinghua.org.cn>
 *   JianFeng Xu        <fenax@video.mdc.tsinghua.edu.cn>
 * \date   : 2003.8
 ************************************************************************
 */
int                                     //  ==> minimum motion cost after search
FastIntegerPelBlockMotionSearch  (pel_t**   orig_pic,     // <--  not used
                  int       ref,          // <--  reference frame (0... or -1 (backward))
                  int       list,
                  int       pic_pix_x,    // <--  absolute x-coordinate of regarded AxB block
                  int       pic_pix_y,    // <--  absolute y-coordinate of regarded AxB block
                  int       blocktype,    // <--  block type (1-16x16 ... 7-4x4)
                  int       pred_mv_x,    // <--  motion vector predictor (x) in sub-pel units MV_pred_space 中值预测矢量
                  int       pred_mv_y,    // <--  motion vector predictor (y) in sub-pel units
                  int*      mv_x,         /* --> motion vector (x) - in pel units
                                      按照H.264标准算法进行的运动矢量预测得到MV_pred
                                      指的是SetMotionVectorPreditor函数预测的MV
                                      和中值预测的区别在于SetMotionVectorPreditor函数预测的MV的参考邻块和当前块必须参
                                      考同一个参考帧，而中值预测的邻块则没有这个要求，二者可能一样，也可能不同*/
                  int*      mv_y,         //  --> motion vector (y) - in pel units
                  int       search_range, // <--  1-d search range in pel units
                  int       min_mcost,    // <--  minimum motion cost (cost for center or huge value)
                  double    lambda)       // <--  lagrangian parameter for determining motion cost
{
  static int Diamond_x[4] = {-1, 0, 1, 0};//对应不同算法  菱形插值
  static int Diamond_y[4] = {0, 1, 0, -1};
  static int Hexagon_x[6] = {2, 1, -1, -2, -1, 1};//六角形插值
  static int Hexagon_y[6] = {0, -2, -2, 0,  2, 2};
  static int Big_Hexagon_x[16] = {0,-2, -4,-4,-4, -4, -4, -2,  0,  2,  4,  4, 4, 4, 4, 2};
  static int Big_Hexagon_y[16] = {4, 3, 2,  1, 0, -1, -2, -3, -4, -3, -2, -1, 0, 1, 2, 3};//大六角形插值

  int   pos, cand_x, cand_y,  mcost;
  pel_t *(*get_ref_line)(int, pel_t*, int, int, int, int);
  int   list_offset   = ((img->MbaffFrameFlag)&&(img->mb_data[img->current_mb_nr].mb_field))? img->current_mb_nr%2 ? 4 : 2 : 0;
  pel_t*  ref_pic       = listX[list+list_offset][ref]->imgY_11;//img->type==B_IMG? Refbuf11 [ref+((mref==mref_fld)) +1] : Refbuf11[ref];
  int   best_pos      = 0;                                        // position with minimum motion cost
  int   max_pos       = (2*search_range+1)*(2*search_range+1);    // number of search positions
  int   lambda_factor = LAMBDA_FACTOR (lambda);                   // factor for determining lagragian motion cost
  int   mvshift       = 2;                  // motion vector shift for getting sub-pel units
  int   blocksize_y   = input->blc_size[blocktype][1];            // vertical block size
  int   blocksize_x   = input->blc_size[blocktype][0];            // horizontal block size
  int   blocksize_x4  = blocksize_x >> 2;                         // horizontal block size in 4-pel units
  int   pred_x        = (pic_pix_x << mvshift) + pred_mv_x;       // predicted position x (in sub-pel units)
  int   pred_y        = (pic_pix_y << mvshift) + pred_mv_y;       // predicted position y (in sub-pel units)
  int   center_x      = pic_pix_x + *mv_x;                        // center position x (in pel units)
  int   center_y      = pic_pix_y + *mv_y;                        // center position y (in pel units)
  int    best_x, best_y;
  int   check_for_00  = (blocktype==1 && !input->rdopt && img->type!=B_SLICE && ref==0);
  int   search_step,iYMinNow, iXMinNow;
  int   i,m, iSADLayer;
  int   iAbort;
  int       N_Bframe = input->successive_Bframe;
  float betaSec,betaThird;
  int height=((img->MbaffFrameFlag)&&(img->mb_data[img->current_mb_nr].mb_field))?img->height/2:img->height;

  //===== set function for getting reference picture lines =====
  if ((center_x > search_range) && (center_x < img->width -1-search_range-blocksize_x) &&
    (center_y > search_range) && (center_y < height-1-search_range-blocksize_y)   )
  {
    get_ref_line = FastLineX;
  }
  else
  {
    get_ref_line = UMVLineX;  //无运动矢量限制，需像素拓展
  }

  //////allocate memory for search state//////////////////////////
  //初始化搜索标记
  memset(McostState[0],0,(2*search_range+1)*(2*search_range+1)*4);

   ///////Threshold defined for early termination///////////////////
  //为早期终止设定门限值
  if(ref>0)
  {
    if(pred_SAD_ref!=0)
    {
      betaSec = Bsize[blocktype]/(pred_SAD_ref*pred_SAD_ref)-AlphaSec[blocktype];
      betaThird = Bsize[blocktype]/(pred_SAD_ref*pred_SAD_ref)-AlphaThird[blocktype];
    }
    else
    {
      betaSec = 0;
      betaThird = 0;
    }
  }
  else
  {
    if(blocktype==1)
    {
      if(pred_SAD_space !=0)
      {
        betaSec = Bsize[blocktype]/(pred_SAD_space*pred_SAD_space)-AlphaSec[blocktype];
        betaThird = Bsize[blocktype]/(pred_SAD_space*pred_SAD_space)-AlphaThird[blocktype];
      }
      else
      {
        betaSec = 0;
        betaThird = 0;
      }
    }
    else
    {
      if(pred_SAD_uplayer !=0)
      {
        betaSec = Bsize[blocktype]/(pred_SAD_uplayer*pred_SAD_uplayer)-AlphaSec[blocktype];
        betaThird = Bsize[blocktype]/(pred_SAD_uplayer*pred_SAD_uplayer)-AlphaThird[blocktype];
      }
      else
      {
        betaSec = 0;
        betaThird = 0;
      }
    }
  }
  /*********检测中值预测矢量**************//*其实就是把得到的mv_pred取整得到的预测矢量*/
  //  MV_pred_space 中值预测矢量
  //check the center median predictor
  cand_x = center_x ;
  cand_y = center_y ;
  mcost = MV_COST (lambda_factor, mvshift, cand_x, cand_y, pred_x, pred_y);//通过计算候选mv所占用的bit得到mv_cost = lambda * bit_of_mv
  mcost = PartCalMad(ref_pic, orig_pic, get_ref_line,blocksize_y,blocksize_x,blocksize_x4,mcost,min_mcost,cand_x,cand_y);//cost = mv_cost + SAD
  McostState[search_range][search_range] = mcost;
  if (mcost < min_mcost)
  {
    min_mcost = mcost;
    best_x = cand_x;
    best_y = cand_y;
  }

  iXMinNow = best_x;
  iYMinNow = best_y;
  for (m = 0; m < 4; m++) //小菱形检测
  {
    cand_x = iXMinNow + Diamond_x[m];
    cand_y = iYMinNow + Diamond_y[m];
    SEARCH_ONE_PIXEL
  }
/*****************原点检测***************************************/
  if(center_x != pic_pix_x || center_y != pic_pix_y)
  {
    cand_x = pic_pix_x ;
    cand_y = pic_pix_y ;
    SEARCH_ONE_PIXEL

    iXMinNow = best_x;
    iYMinNow = best_y;
    for (m = 0; m < 4; m++)//小菱形检测
    {
      cand_x = iXMinNow + Diamond_x[m];
      cand_y = iYMinNow + Diamond_y[m];
      SEARCH_ONE_PIXEL
    }
  }
 /**********************上层块预测矢量检测*********************************/
    if(blocktype>1)//
  {
    cand_x = pic_pix_x + (pred_MV_uplayer[0]/4);
    cand_y = pic_pix_y + (pred_MV_uplayer[1]/4);
    SEARCH_ONE_PIXEL
    if ((min_mcost-pred_SAD_uplayer)<pred_SAD_uplayer*betaThird)
      goto third_step;
    else if((min_mcost-pred_SAD_uplayer)<pred_SAD_uplayer*betaSec)
      goto sec_step;
  }
    /****************相应块预测***************************************/

  //coordinate position prediction
  if ((img->number > 1 + ref && ref!=-1) || (list == 1 && (Bframe_ctr%N_Bframe) > 1))  //for debug
  {
    cand_x = pic_pix_x + pred_MV_time[0]/4;
    cand_y = pic_pix_y + pred_MV_time[1]/4;
    SEARCH_ONE_PIXEL
  }
  /******************相邻参考帧预测*********************************/

  //prediciton using mV of last ref moiton vector
  if (input->PicInterlace == FIELD_CODING)//场编码，用最近的场MV预测
  {
    if ((list==0 && ref > 0) || (img->type == B_SLICE && list == 0 && (ref==0 ||ref==2 ) ))
      //Notes: for interlace case, ref==1 should be added
    {
      cand_x = pic_pix_x + pred_MV_ref[0]/4;
      cand_y = pic_pix_y + pred_MV_ref[1]/4;
      SEARCH_ONE_PIXEL
    }
  }
  else
  {   //多参考帧预测时，用另一帧的MV预测
    if ((list==0 && ref > 0) || (img->type == B_SLICE && list == 0 && ref==0 ))
      //Notes: for interlace case, ref==1 should be added
    {
      cand_x = pic_pix_x + pred_MV_ref[0]/4;
      cand_y = pic_pix_y + pred_MV_ref[1]/4;
      SEARCH_ONE_PIXEL
    }
  }
  //small local search
  iXMinNow = best_x;
  iYMinNow = best_y;
  for (m = 0; m < 4; m++)//小菱形搜索
  {
    cand_x = iXMinNow + Diamond_x[m];
    cand_y = iYMinNow + Diamond_y[m];
    SEARCH_ONE_PIXEL
  } 

  //early termination algrithm, refer to JVT-D016
   //根据SAD值判断需要跳转的步骤，SAD较小时转到步骤3，较大时转到步骤2，很大时转到步骤1
    EARLY_TERMINATION

  if(blocktype>6)
    goto sec_step;
  else
    goto first_step;

first_step: //Unsymmetrical-cross search 不甚满意
  iXMinNow = best_x;
  iYMinNow = best_y;

  for(i=1;i<=search_range/2;i++)//水平方向搜索
  {
    search_step = 2*i - 1;
    cand_x = iXMinNow + search_step;
    cand_y = iYMinNow ;
    SEARCH_ONE_PIXEL
    cand_x = iXMinNow - search_step;
    cand_y = iYMinNow ;
    SEARCH_ONE_PIXEL
  }

  //垂直方向搜索，注意垂直方向搜索点比水平方向少，考虑到了水平方向较垂直方向重要
  for(i=1;i<=search_range/4;i++)
  {
    search_step = 2*i - 1;
    cand_x = iXMinNow ;
    cand_y = iYMinNow + search_step;
    SEARCH_ONE_PIXEL
    cand_x = iXMinNow ;
    cand_y = iYMinNow - search_step;
    SEARCH_ONE_PIXEL
  }
  //early termination algrithm, refer to JVT-D016
    //在这里也进行中止、跳转检测，考虑到一般序列中含有大量水平、垂直方向的运动。
    EARLY_TERMINATION

  iXMinNow = best_x;
  iYMinNow = best_y;
    //螺旋搜索，类似全搜索法，只搜索前25点，相当于5×5区域全搜索
  for(pos=1;pos<25;pos++)
  {
    cand_x = iXMinNow + spiral_search_x[pos];
    cand_y = iYMinNow + spiral_search_y[pos];
    SEARCH_ONE_PIXEL
  }
  //early termination algrithm, refer to JVT-D016
    EARLY_TERMINATION

   // Uneven Multi-Hexagon-grid Search
    //超六边形模板搜索，（多圈）
  for(i=1;i<=search_range/4; i++)
  {
    iAbort = 0;
    for (m = 0; m < 16; m++)
    {
      cand_x = iXMinNow + Big_Hexagon_x[m]*i;
      cand_y = iYMinNow + Big_Hexagon_y[m]*i;
      SEARCH_ONE_PIXEL1(1)
    }
    if (iAbort)
    {
      //early termination algrithm, refer to JVT-D016
      EARLY_TERMINATION
    }
  }

// 六边形模板反复搜索（也可以用大菱形代替），搜索完后进入第三步骤
sec_step:  //Extended Hexagon-based Search 满意
      iXMinNow = best_x;
      iYMinNow = best_y;
      for(i=0;i<search_range;i++)
      {
        iAbort = 1;
        for (m = 0; m < 6; m++)
        {
          cand_x = iXMinNow + Hexagon_x[m];
          cand_y = iYMinNow + Hexagon_y[m];
          SEARCH_ONE_PIXEL1(0)
        }
        if(iAbort)
          break;
        iXMinNow = best_x;
        iYMinNow = best_y;
      }
// 小菱形模板反复搜索，得到最终的运动矢量
third_step: // the third step with a small search pattern  很满意
      iXMinNow = best_x;
      iYMinNow = best_y;
      for(i=0;i<search_range;i++)
      {
        iSADLayer = 65536;
        iAbort = 1;
        for (m = 0; m < 4; m++)
        {
          cand_x = iXMinNow + Diamond_x[m];
          cand_y = iYMinNow + Diamond_y[m];
          SEARCH_ONE_PIXEL1(0)
        }
        if(iAbort)
          break;
        iXMinNow = best_x;
        iYMinNow = best_y;
      }

      *mv_x = best_x - pic_pix_x;
      *mv_y = best_y - pic_pix_y;
      return min_mcost;
  }