h.264全搜索以及快速全搜索算法
Full Search
全搜索算法是最简单暴力的一种搜索算法，对搜索范围内的所有像素点都进行匹配对比，选出最合适的运动向量，以下就是一个搜索范围为4的全搜索范围（单个像素点）
/*! *********************************************************************** * \brief按照螺旋搜索顺序进行全搜索 * Full pixel block motion search * 目标是得到(mv_x,mv_y)和min_mcost，(mv_x,mv_y)指示从哪里开始做分像素搜索，search center * 后者用来跟分像素搜索结果做比较 *********************************************************************** */ int // ==> minimum motion cost after search FullPelBlockMotionSearch (pel_t** orig_pic, // <-- original pixel values for the AxB block int ref, // <-- reference frame (0... or -1 (backward)) int list, int pic_pix_x, // <-- absolute x-coordinate of regarded AxB blockAxB宏块原点在图像中的绝对坐标 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, // <--> in: search center (x) / out: motion vector (x) - in pel units int* mv_y, // <--> in: search center (y) / out: 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 { int pos, cand_x, cand_y, y, x4, mcost; pel_t *orig_line, *ref_line; pel_t *(*get_ref_line)(int, pel_t*, int, int, int, int);// //参考帧偏移量 帧场自适应且宏块地址为偶数=4 帧场自适应宏块地址为奇数=2 非帧场自适应=0 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; int img_width = listX[list+list_offset][ref]->size_x; int img_height = listX[list+list_offset][ref]->size_y; int best_pos = 0; // position with minimum motion cost //计算最大需要搜索的位置个数 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 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 << 2) + pred_mv_x; // predicted position x (in sub-pel units)1/4子像素为单位的预测MV int pred_y = (pic_pix_y << 2) + 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 check_for_00 = (blocktype==1 && !input->rdopt && img->type!=B_SLICE && ref==0); //===== set function for getting reference picture lines ===== //通过判断搜索范围会不会出界，设置获取参考像素值的函数 if ((center_x > search_range) && (center_x < img->width -1-search_range-blocksize_x) && (center_y > search_range) && (center_y < img->height-1-search_range-blocksize_y) ) { get_ref_line = FastLineX;//未出界 } else { get_ref_line = UMVLineX;//出界 } //===== loop over all search positions ===== //max_pos是搜索位置的个数,计算见上面 for (pos=0; pos<max_pos; pos++) { //--- set candidate position (absolute position in pel units) --- /*(center_x,center_y)是由预测MV估计出来的搜索中心，在以它为中心的范围内， 对按照螺旋形顺序排列的候选点进行搜索， 每个候选点都是一个可能参考块的左上角起始点 */ cand_x = center_x + spiral_search_x[pos];//螺旋搜索 cand_y = center_y + spiral_search_y[pos]; //--- initialize motion cost (cost for motion vector) and check --- //计算MVD的代价，换算成四分之一像素(cand--candidate候选点) mcost = MV_COST (lambda_factor, 2, cand_x, cand_y, pred_x, pred_y); if (check_for_00 && cand_x==pic_pix_x && cand_y==pic_pix_y) {//螺旋搜索到的点为原点,不过为什么是减去16bit? mcost -= WEIGHTED_COST (lambda_factor, 16); } //如果只是MV的代价就已经大于现有的最小代价就舍弃 if (mcost >= min_mcost) continue; //--- add residual cost to motion cost --- //blocksize_y blocksize_x4 是分块大小16x16 16x8 8x16...... for (y=0; y<blocksize_y; y++) { //(cand_x,cand_y+y)是一行的起始坐标，y++ 遍历每一行 ref_line = get_ref_line (blocksize_x, ref_pic, cand_y+y, cand_x, img_height, img_width); orig_line = orig_pic [y]; //计算当前帧和参考帧的像素残差 for (x4=0; x4<blocksize_x4; x4++) //以4个为一组计算 { mcost += byte_abs[ *orig_line++ - *ref_line++ ]; mcost += byte_abs[ *orig_line++ - *ref_line++ ]; mcost += byte_abs[ *orig_line++ - *ref_line++ ]; mcost += byte_abs[ *orig_line++ - *ref_line++ ]; } if (mcost >= min_mcost) //如果已经比最小代价大，就没必要计算下面的行了 { break; } } //--- check if motion cost is less than minimum cost --- //记录下最小代价和最佳匹配位置 if (mcost < min_mcost) { best_pos = pos; min_mcost = mcost; } } //===== set best motion vector and return minimum motion cost ===== if (best_pos) { *mv_x += spiral_search_x[best_pos]; //因为螺旋搜索数组中记录的是该位置的点 *mv_y += spiral_search_y[best_pos]; //与(center_x,center_y)的差 } return min_mcost; //返回最小代价 }
//螺旋搜索（全搜索）位置初始化 for (k=1, l=1; l<=max(1,search_range); l++) { for (i=-l+1; i< l; i++) { spiral_search_x[k] = i; spiral_search_y[k++] = -l; spiral_search_x[k] = i; spiral_search_y[k++] = l; /* * * 9 3 5 7 10 * 1 0 2 11 1 0 2 12 * 13 4 6 8 14 * */ } for (i=-l; i<=l; i++) { spiral_search_x[k] = -l; spiral_search_y[k++] = i; spiral_search_x[k] = l; spiral_search_y[k++] = i; /* 15 17 19 21 23 * 3 5 7 9 3 5 7 10 * 1 0 2 11 1 0 2 12 * 4 6 8 13 4 6 8 14 * 16 18 20 22 24 */ } }
Fast Full Search
由于运动搜索时有多种块的类型（16x16,8x16,8x8,4x4等）因此，在全搜索时，会由于位置重叠而产生同一处的像素残差多次计算的情况，为了避免这种情况，可以一次性把搜索范围内的所有像素残差计算出来，不同块类型只需要把残差进行组合即可得到该类型的SAD
/*! *********************************************************************** * \brief快速正像素搜索 * Fast Full pixel block motion search * 目标是得到(mv_x,mv_y)和min_mcost，(mv_x,mv_y)指示从哪里开始做分像素搜索，search center * 后者用来跟分像素搜索结果做比较 *********************************************************************** */ int // ==> minimum motion cost after search FastFullPelBlockMotionSearch (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 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 { int pos, offset_x, offset_y, cand_x, cand_y, mcost; 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 best_pos = 0; // position with minimum motion cost int block_index; // block index for indexing SAD array int* block_sad; // pointer to SAD array block_index = (pic_pix_y-img->opix_y)+((pic_pix_x-img->opix_x)>>2); // block index for indexing SAD array block_sad = BlockSAD[list][ref][blocktype][block_index]; // pointer to SAD array //===== set up fast full integer search if needed / set search center ===== if (!search_setup_done[list][ref])//对一个参考帧只做一次 { //计算搜索范围所有位置所有分块模式的SAD(整像素) SetupFastFullPelSearch (ref, list); } offset_x = search_center_x[list][ref] - img->opix_x; //搜索中心相对原宏块的偏移 offset_y = search_center_y[list][ref] - img->opix_y; //===== cost for (0,0)-vector: it is done before, because MVCost can be negative ===== //原点(这里的原点都是是当前块所在的位置) if (!input->rdopt) { //把刚才计算的SAD 跟mv代价相加得到总代价 mcost = block_sad[pos_00[list][ref]] + MV_COST (lambda_factor, 2, 0, 0, pred_mv_x, pred_mv_y); if (mcost < min_mcost) { min_mcost = mcost; best_pos = pos_00[list][ref];//每帧搜索中心的位置 } } //===== loop over all search positions ===== for (pos=0; pos<max_pos; pos++, block_sad++) { //--- check residual cost --- if (*block_sad < min_mcost) { //--- get motion vector cost --- //计算出搜索位置，按照螺旋形顺序spiral_search_xy cand_x = offset_x + spiral_search_x[pos]; cand_y = offset_y + spiral_search_y[pos]; mcost = *block_sad;//在SetupFastFullPelSearch已经做好 mcost += MV_COST (lambda_factor, 2, cand_x, cand_y, pred_mv_x, pred_mv_y); //计算 MV 代价 //--- check motion cost --- if (mcost < min_mcost) { min_mcost = mcost; best_pos = pos; } } } //===== set best motion vector and return minimum motion cost ===== *mv_x = offset_x + spiral_search_x[best_pos];//根据代价最小，计算出最佳MV *mv_y = offset_y + spiral_search_y[best_pos]; return min_mcost; }
Edge Process
通常来说，计算SAD是以一行一行为单位进行，不过在进行搜索时，难免会有运动向量指向图像外的区域，图像以外的这些区域的像素取值为图像边界的值，即
$Pic[x,y]=\left\{\begin{matrix}
Pic[0,y] & x<0\\
Pic[width-1,y] & x\geq width\\
Pic[x,0] & 0\leq x < width,y<0\\
Pic[x,height-1] & 0\leq x < width,y \geq height\\
Pic[x,y] & other
\end{matrix}\right.$
/*如果参考块超出参考帧边界，用边界值进行填充*/ pel_t *UMVLineX (int size, pel_t* Pic, int y, int x, int height, int width) { int i, maxx; pel_t *Picy; Picy = Pic + max(0,min(height-1,y)) * width; //先把y范围限制在(0,height-1)内 if (x < 0) // Left edge { maxx = min(0,x+size); //搜索范围可以大于16的，所以x+16是可以小于0的 for (i = x; i < maxx; i++) //把出界的部分都赋值为边界上的值，一画图就理解了 { line[i-x] = Picy [0]; // Replicate left edge pixel } maxx = x+size; //把没出界的像素也拷贝到line数组中 for (i = 0; i < maxx; i++) // Copy non-edge pixels line[i-x] = Picy [i]; } else if (x > width-size) // Right edge 同理 { maxx = width; for (i = x; i < maxx; i++) { line[i-x] = Picy [i]; // Copy non-edge pixels } maxx = x+size; for (i = max(width,x); i < maxx; i++) { line[i-x] = Picy [width-1]; // Replicate right edge pixel } } else // No edge ，则返回y行x列的地址 { return Picy + x; } return line; //否则，返回line数组的地址 }
