OpenCV Haar AdaBoost源代码改进（比EMCV快6倍）
这几天研究了OpenCV源代码 Haar AdaBoost算法，作了一下改进
1.去掉了全部动态分配内存的操作。对嵌入式系统有一定的速度提升
2.凝视覆盖了大量关键代码
3.降低了代码一半的体积，而且降低了部分健壮性的代码，速度比OpenCV源代码提升16%
4.改动了大量数据结构，不依赖CV源代码直接编译
5.去掉了double型，改成Int
6.开方改成查表
7.除法改成乘法加位移
注：使用时请注意，现仅支持单分支的Stages和单结点的Classifier训练好的结果集
在720MHZ的DSP板子上对一幅352*288的灰度图像进行人脸检測仅仅需300ms，比EMCV快6倍
完整PC版project链接 （VC6.0能直接编译。但没有5。6，7步的优化）点击打开链接
完整DSP版project链接 （CCS3.0能直接编译，包括全部优化）点击打开链接
DSP优化的关键代码实比例如以下（这个版本号在CCS下编译，若想用VC6.0直接编译，还要改动一定的数据结构）
Haar.cpp
<pre name="code" class="cpp">#include "Haar.h"
#include "loadCascade.h"
#include "Util.h"
#include "stdio.h"
#include "string.h"
#include <math.h>
#include <stdint.h>
#include <c6x.h> /*******************Global************************************/
HaarClassifierCascade *cascade ;
//HidHaarClassifierCascade hid_cascade;
//32bits cell Mat
int MatPool32[MaxMatNum][MAXROWS][MAXCOLS];
//8bits cell
unsigned char MatPool8[MaxMatNum][MAXROWS][MAXCOLS]; //8bits*3 cell
unsigned char ImgRGBPool8[MaxMatNum][RGBCHANNEL][MAXROWS][MAXCOLS]; //64bits cell
_int64 MatPool64[MaxMatNum][MAXROWS][MAXCOLS]; //候选区域坐标节点并查集
PTreeNode PTreeNodes[MAXPTREENODES]; char HidCascade[MAXHIDCASCADE]; //分类器检測结果区域序列
Sequence result_seq; //==================================================================
//函数名： IsEqual
//作者： qiurenbo
//日期： 2014-10-1
//功能： 推断两个矩形是否邻接
//输入參数：_r1 _r2 候选区域矩形
//返回值： 返回类似性（是否是邻接的矩形）
//改动记录：
//==================================================================
int IsEqual( const void* _r1, const void* _r2)
{
const Rect* r1 = (const Rect*)_r1;
const Rect* r2 = (const Rect*)_r2;
int distance5x = r1->width ;//int distance = cvRound(r1->width*0.2); return r2->x*5 <= r1->x*5 + distance5x &&
r2->x*5 >= r1->x*5 - distance5x &&
r2->y*5 <= r1->y*5 + distance5x &&
r2->y*5 >= r1->y*5 - distance5x &&
r2->width*5 <= r1->width * 6 &&
r2->width * 6 >= r1->width*5;
} //==================================================================
//函数名： ReadFaceCascade
//作者： qiurenbo
//日期： 2014-10-1
//功能： 依据候选区域的类似性（IsEqual函数）建立并查集
//输入參数：seq 候选目标区域序列
//返回值： 返回分类后的类别数
//改动记录：
//==================================================================
int SeqPartition( const Sequence* seq )
{
Sequence* result = 0;
//CvMemStorage* temp_storage = 0;
int class_idx = 0; memset(PTreeNodes, 0, MAXPTREENODES*sizeof(PTreeNode)); int i, j; //建立以seq中元素为根节点的森林
for( i = 0; i < seq->total; i++ )
PTreeNodes[i].element = (char*)&seq->rectQueue[i]; //遍历全部根节点
for( i = 0; i < seq->total; i++ )
{
PTreeNode* node = &PTreeNodes[i];
PTreeNode* root = node;
//确保node中元素指针不为空
if( !node->element )
continue; //找到元素在树中的根结点
while( root->parent )
root = root->parent; for( j = 0; j < seq->total; j++ )
{
PTreeNode* node2 = &PTreeNodes[j]; //确保1.node中元素指针不为空
// 2.且不是同一个node结点
// 3.且是类似区域
// 若是类似区域，则合并元素
if( node2->element && node2 != node &&
IsEqual( node->element, node2->element))
{
PTreeNode* root2 = node2; //找到元素在树中的根结点
while( root2->parent )
root2 = root2->parent; //合并的前提是不在一颗树中
if( root2 != root )
{
//秩小的树归入秩大的树中
if( root->rank > root2->rank )
root2->parent = root;
//秩相等的时候才改变树的秩
else
{
root->parent = root2;
root2->rank += root->rank == root2->rank;
root = root2;
}
//assert( root->parent == 0 ); // 路径压缩。子节点node2直接指向根节点
while( node2->parent )
{
PTreeNode* temp = node2;
node2 = node2->parent;
temp->parent = root;
} // 路径压缩，子节点node直接指向根节点
node2 = node;
while( node2->parent )
{
PTreeNode* temp = node2;
node2 = node2->parent;
temp->parent = root;
}
}
} }
} for( i = 0; i < seq->total; i++ )
{
PTreeNode* node = &PTreeNodes[i];
int idx = -1; if( node->element )
{
while( node->parent )
node = node->parent; //计算有几棵并查树，巧妙地利用取反避免反复计算
if( node->rank >= 0 )
node->rank = ~class_idx++;
idx = ~node->rank;
} } return class_idx;
} //==================================================================
//函数名： ReadFaceCascade
//作者： qiurenbo
//日期： 2014-09-30
//功能： 读取Cascade文件
//输入參数：void
//返回值： void
//改动记录：
//==================================================================
void ReadFaceCascade()
{
int i;
//load cascade
cascade = (HaarClassifierCascade*)HaarClassifierCascade_face; //load stages
int stage_size = StageClassifier_face[0];
HaarStageClassifier *stages ;
stages = (HaarStageClassifier *)(StageClassifier_face+1); //load classifier
int classifier_size = Classifier_face[0];
HaarClassifier *cls ;
cls = (HaarClassifier*) (Classifier_face+1); int class_info_size = class_info[0];
int * cls_info ;
cls_info = (int*)(class_info+1); //link cascade with stages
cascade->stage_classifier = stages;
//link stages。classifiers
int offset=0;
int offset_t=(sizeof(HaarFeature)/sizeof(int));
int offset_l=offset_t+1;
int offset_r=offset_t+2;
int offset_a=offset_t+3;
int offset_total=0;
for(i=0;i<stage_size;++i)
{
(stages+i)->classifier = (cls+offset);
offset +=(stages+i)->count;
} offset_total = 5+ (sizeof(HaarFeature)/sizeof(int));
//link classifiers and haar_featrue;
for(i=0;i<classifier_size;++i)
{
HaarClassifier *cs= cls+i; cs->haar_feature = (HaarFeature*)(cls_info+i*offset_total);
cs->threshold = (int*)(cls_info+i*offset_total+offset_t);
cs->left =(int*)(cls_info+i*offset_total+offset_l);
cs->right=(int*)(cls_info+i*offset_total+offset_r);
cs->alpha=(int*)(cls_info+i*offset_total+offset_a); }
}
//==================================================================
//函数名： IntegralImage
//作者： qiurenbo
//日期： 2014-09-26
//功能： 从矩阵池中获取rows * cols的矩阵
//输入參数：mat 矩阵结构体地址
// rows 待分配的行数
// cols 待分配的列数
// type 待分配的矩阵类型
// matIndex 从矩阵池中分配的矩阵序列（手动指定..）
//返回值： void
//改动记录：
//================================================================== void GetMat(void* mat, int rows, int cols, int type, int matIndex)
{
switch(type)
{
case BITS8:
((Mat8*)mat)->rows = rows;
((Mat8*)mat)->cols = cols;
((Mat8*)mat)->mat8Ptr = (Mat8Ptr)&MatPool8[matIndex];
break; case BITS32:
((Mat32*)mat)->rows = rows;
((Mat32*)mat)->cols = cols;
((Mat32*)mat)->mat32Ptr = (Mat32Ptr)&MatPool32[matIndex];
break; case BITS64:
((Mat64*)mat)->rows = rows;
((Mat64*)mat)->cols = cols;
((Mat64*)mat)->mat64Ptr = (Mat64Ptr)&MatPool64[matIndex];
break;
}
} //==================================================================
//函数名： IntegralImage
//作者： qiurenbo
//日期： 2014-09-26
//功能： 计算目标检測区域的积分图
//输入參数：src 待检測目标所在矩阵起始
// srcstep 待检測区域列数
// sum 积分图矩阵 (W+1)*(H+1)
// sumstep 积分图矩阵列数
// sqsum 平方和图矩阵 (W+1)*(H+1)
// sqsumstep 平方和图矩阵列数
// size 待检測区域大小 W*H
//
//
//返回值： void
//改动记录：
//==================================================================
void IntegralImage(ImgPtr src, int srcstep,
Mat32Ptr sum, int sumstep,
Mat64Ptr sqsum, int sqsumstep,
Size size)
{
int s = 0;
_int64 sq = 0;
//移动指针到积分图的下一行,第一行全为0
sum += sumstep + 1;
sqsum += sqsumstep + 1; //y代表相对于输入检測矩阵起始第几行
for(int y = 0; y < size.height; y++, src += srcstep,
sum += sumstep, sqsum += sqsumstep )
{
//sum和sqsum为(W+1)*(H+1)大小矩阵，故将第一列置为0
sum[-1] = 0;
sqsum[-1] = 0; for(int x = 0 ; x < size.width; x++ )
{
int it = src[x];
int t = (it); //查表计算平方
_int64 tq = CV_8TO16U_SQR(it);
//s代表行上的累加和
s += t;
//sq代表行上的累加和
sq += tq;
t = sum[x - sumstep] + s;
tq = sqsum[x - sqsumstep] + sq;
sum[x] = t;
sqsum[x] = (_int64)tq;
}
}
} //==================================================================
//函数名： Integral
//作者： qiurenbo
//日期： 2014-09-26
//功能： 计算目标检測区域的积分图
//输入參数：image 图像
// sumImage 积分图指针
// sumSqImage 平方和图指针
//返回值： void
//改动记录：
//==================================================================
void Integral(Image* image, Mat32* sumImage, Mat64* sumSqImage)
{ //取保地址空间已经分配，从数组中
if (image == NULL || sumImage == NULL || sumSqImage == NULL)
return; Image*src = (Image*)image;
Mat32 *sum = (Mat32*)sumImage;
Mat64 *sqsum = (Mat64*)sumSqImage; Size size;
size.height = src->rows;
size.width = src->cols; IntegralImage(src->imgPtr, src->cols,
sum->mat32Ptr, sum->cols,
sqsum->mat64Ptr, sqsum->cols,size); }
//==================================================================
//函数名： AlignPtr
//作者： qiurenbo
//日期： 2014-10-03
//功能： 按algin字节对齐
//输入參数：ptr 要对齐的指针
// align 对齐的字节数
//返回值： void*
//改动记录：
//==================================================================
void* AlignPtr( const void* ptr, int align)
{ return (void*)( ((unsigned int)ptr + align - 1) & ~(align-1) );
}
//==================================================================
//函数名： CreateHidHaarClassifierCascade
//作者： qiurenbo
//日期： 2014-09-28
//功能： 创建隐式积分图加快计算速度
//输入參数：cascade 级联分类器指针
//返回值： static HidHaarClassifierCascade* 返回一个隐式级联分类器指针
//改动记录：
//==================================================================
static HidHaarClassifierCascade*
CreateHidHaarClassifierCascade(HaarClassifierCascade* cascade)
{ cascade->hid_cascade = (struct HidHaarClassifierCascade *)HidCascade;
//分配栈空间
HidHaarClassifierCascade* out = (struct HidHaarClassifierCascade *)HidCascade;
const int icv_stage_threshold_bias = 419; //0.0001*(2^22)=419.4304 HidHaarClassifier* haar_classifier_ptr;
HidHaarTreeNode* haar_node_ptr;
int i, j, l; int total_classifiers = 2135;
int total_nodes = 0; int has_tilted_features = 0;
int max_count = 0; /* 初始化HidCascade头 */
out->count = cascade->count;
out->stage_classifier = (HidHaarStageClassifier*)(out + 1);
//out->stage_classifier = (HidHaarStageClassifier*)AlignPtr(out + 1, 4);
//classifier起始地址
haar_classifier_ptr = (HidHaarClassifier*)(out->stage_classifier + cascade->count);
//haar_classifier_ptr = (HidHaarClassifier*)AlignPtr(out->stage_classifier + cascade->count, 4);
//node起始地址
//haar_node_ptr = (HidHaarTreeNode*)AlignPtr(haar_classifier_ptr + total_classifiers, 4);
haar_node_ptr = (HidHaarTreeNode*)(haar_classifier_ptr + total_classifiers);
out->is_stump_based = 1;
out->is_tree = 0; // 用cascade初始化HidCascade
for( i = 0; i < cascade->count; i++ )
{ //用cascades Stage初始化HidCascade的Stage
HaarStageClassifier* stage_classifier = cascade->stage_classifier + i;
HidHaarStageClassifier* hid_stage_classifier = out->stage_classifier + i; hid_stage_classifier->count = stage_classifier->count;
hid_stage_classifier->threshold = stage_classifier->threshold - icv_stage_threshold_bias;
//hid_stage_classifier->classifier = (struct HidHaarClassifier *)&HidClassifiers[i];
hid_stage_classifier->classifier = haar_classifier_ptr;
//初始化为二特征，以下会依据真实的特征数至1或0（三特征）
hid_stage_classifier->two_rects = 1;
haar_classifier_ptr += stage_classifier->count; //Stage构成一颗退化的二叉树（单分支），每一个结点最多仅仅有一个孩子
hid_stage_classifier->parent = (stage_classifier->parent == -1)
? NULL : out->stage_classifier + stage_classifier->parent;
hid_stage_classifier->next = (stage_classifier->next == -1)
? NULL : out->stage_classifier + stage_classifier->next;
hid_stage_classifier->child = (stage_classifier->child == -1)
? NULL : out->stage_classifier + stage_classifier->child ; //推断该stage是否为树状结构（多分枝）
out->is_tree |= hid_stage_classifier->next != NULL; //赋值classifer属性
for( j = 0; j < stage_classifier->count; j++ )
{
HaarClassifier* classifier = stage_classifier->classifier + j;
HidHaarClassifier* hid_classifier = hid_stage_classifier->classifier + j;
int node_count = classifier->count; int* alpha_ptr = (int*)(haar_node_ptr + node_count); hid_classifier->count = node_count;
hid_classifier->node = haar_node_ptr;
hid_classifier->alpha = alpha_ptr; //赋值node属性
for( l = 0; l < node_count; l++ )
{
HidHaarTreeNode* node = hid_classifier->node + l;
HaarFeature* feature = classifier->haar_feature + l;
memset( node, -1, sizeof(*node) );
node->threshold = classifier->threshold[l];
node->left = classifier->left[l];
node->right = classifier->right[l]; //对特征数目进行推断,若是三特征，则至two_rects为0
if( (feature->rect[2].weight) == 0 ||
feature->rect[2].r.width == 0 ||
feature->rect[2].r.height == 0 )
memset( &(node->feature.rect[2]), 0, sizeof(node->feature.rect[2]) );
else
hid_stage_classifier->two_rects = 0;
} //赋值alpha
memcpy( hid_classifier->alpha, classifier->alpha, (node_count+1)*sizeof(hid_classifier->alpha[0]));
haar_node_ptr = (HidHaarTreeNode*)(alpha_ptr+node_count + 1); //推断cascade中的分类器是否是树桩分类器，仅仅有根结点的决策树
out->is_stump_based &= node_count == 1;
}
} //cascade->hid_cascade = out;
//assert( (char*)haar_node_ptr - (char*)out <= datasize ); return out;
} //==================================================================
//函数名： SetImagesForHaarClassifierCascade
//作者： qiurenbo
//日期： 2014-09-29
//功能： 依据尺度调整Haar特征的大小和权重
//输入參数：cascade 级联分类器指针
// sum 积分图
// sqsum 平方和积分图
// scale32x 尺度
//返回值： 无
//改动记录：
//==================================================================
void SetImagesForHaarClassifierCascade(HaarClassifierCascade* _cascade, Mat32* sum, Mat64* sqsum, int scale32x)
{ HidHaarClassifierCascade* hidCascade;
int coi0 = 0, coi1 = 0;
int i;
Rect equ_rect;
int weight_scale;
HaarFeature* feature;
HidHaarFeature* hidfeature;
int sum0 = 0, area0 = 0;
Rect r[3];
Rect tr;
int correction_ratio; //依据尺度获取窗体大小
_cascade->scale32x = scale32x;
_cascade->real_window_size.width = (_cascade->orig_window_size.width * scale32x + 16)>>5 ;
_cascade->real_window_size.height = (_cascade->orig_window_size.height * scale32x +16) >> 5; //设置隐式级联分类器的积分图
hidCascade = _cascade->hid_cascade;
hidCascade->sum = sum;
hidCascade->sqsum = sqsum; //依据尺度设置积分图起始矩阵的位置
equ_rect.x = equ_rect.y = (scale32x+16)>>5;
equ_rect.width = ((_cascade->orig_window_size.width-2)*scale32x + 16 ) >> 5; //+0.5是为了四舍五入
equ_rect.height = ((_cascade->orig_window_size.height-2)*scale32x + 16 ) >> 5;
weight_scale = equ_rect.width*equ_rect.height;
hidCascade->window_area = weight_scale; //矩形面积 //获取积分图上起始矩阵四个像素的坐标
hidCascade->p0 = sum->mat32Ptr + (equ_rect.y) * sum->cols+ equ_rect.x;
hidCascade->p1 = sum->mat32Ptr + (equ_rect.y) * sum->cols + equ_rect.x + equ_rect.width;
hidCascade->p2 = sum->mat32Ptr + (equ_rect.y + equ_rect.height) * sum->cols + equ_rect.x;
hidCascade->p3 = sum->mat32Ptr + (equ_rect.y + equ_rect.height) * sum->cols + equ_rect.x + equ_rect.width; //获取平方和积分图上起始矩阵四个像素的坐标
hidCascade->pq0 = sqsum->mat64Ptr + (equ_rect.y) * sqsum->cols+ equ_rect.x;
hidCascade->pq1 = sqsum->mat64Ptr + (equ_rect.y) * sqsum->cols+ equ_rect.x + equ_rect.width;
hidCascade->pq2 = sqsum->mat64Ptr + (equ_rect.y + equ_rect.height) * sqsum->cols+ equ_rect.x;
hidCascade->pq3 = sqsum->mat64Ptr + (equ_rect.y + equ_rect.height) * sqsum->cols+ equ_rect.x + equ_rect.width; //遍历每一个Classifer所使用的特征，对它们进行尺度放大，并将改变的值赋给HidCascade，隐式级联分类器
for( i = 0; i < hidCascade->count; i++ )
{
int j, k, l;
for( j = 0; j < hidCascade->stage_classifier[i].count; j++ )
{
for( l = 0; l < hidCascade->stage_classifier[i].classifier[j].count; l++ )
{
feature = &_cascade->stage_classifier[i].classifier[j].haar_feature[l]; hidfeature = &hidCascade->stage_classifier[i].classifier[j].node[l].feature;
sum0 = 0;
area0 = 0; for( k = 0; k < CV_HAAR_FEATURE_MAX; k++ )
{
if( !hidfeature->rect[k].p0 )
break; r[k] = feature->rect[k].r; //左上角坐标和矩阵长宽都按尺度放大
tr.x = (r[k].x * scale32x + 16) >> 5;
tr.width = (r[k].width * scale32x + 16) >> 5;
tr.y = ( r[k].y * scale32x + 16 ) >> 5;
tr.height = ( r[k].height * scale32x +16 ) >> 5; correction_ratio = weight_scale; //设置矩阵四个顶点在积分图中的位置（为了计算特征方便）
hidfeature->rect[k].p0 = sum->mat32Ptr + tr.y * sum->cols + tr.x;
hidfeature->rect[k].p1 = sum->mat32Ptr + tr.y * sum->cols + tr.x + tr.width;
hidfeature->rect[k].p2 = sum->mat32Ptr + (tr.y + tr.height) *sum->cols + tr.x;
hidfeature->rect[k].p3 = sum->mat32Ptr + (tr.y + tr.height) *sum->cols + tr.x + tr.width; //rect[1] = weight/area， 左移22位是为了避免浮点计算，将权值/检測窗体面积（不断扩大），减少权值
hidfeature->rect[k].weight = ((feature->rect[k].weight)<< NODE_THRESHOLD_SHIFT)/(correction_ratio); if( k == 0 )
area0 = tr.width * tr.height;
else
sum0 += hidfeature->rect[k].weight * tr.width * tr.height; }
//rect[0].weight ,权重和特征矩形面积成反比
hidfeature->rect[0].weight = (int)(-sum0/area0); } /* l */
} /* j */
} }; uint64_t block1 = 0;
//uint64_t block2 = 0;
//==================================================================
//函数名： RunHaarClassifierCascade
//作者： qiurenbo
//日期： 2014-09-30
//功能： 在指定窗体范围计算特征
//输入參数：_cascade 级联分类器指针
// pt 检測窗体左上角坐标
// start_stage 起始stage下标
//返回值： <=0 未检測到目标或參数有问题
// 1 成功检測到目标
//改动记录：
//====================================================================
int RunHaarClassifierCascade( HaarClassifierCascade* _cascade, Point& pt, int start_stage )
{ int result = -1; int p_offset, pq_offset;
int i, j;
_int64 rectsum, variance_factor;
int variance_norm_factor;
HidHaarClassifier* classifier;
HidHaarTreeNode* node;
int sum, t, a, b;
int stage_sum; /* uint64_t start_time, end_time, overhead, cyclecountSet=0, cyclecountRun=0;
//In the initialization portion of the code:
TSCL = 0; //enable TSC
start_time = _itoll(TSCH, TSCL);
end_time = _itoll(TSCH, TSCL);
overhead = end_time-start_time; //Calculating the overhead of the method.*/
HidHaarClassifierCascade* hidCascade; if (_cascade == NULL)
return -1; hidCascade = _cascade->hid_cascade;
if( !hidCascade )
return -1; //确保矩形的有效性，并防止计算窗体出边界
if( pt.x < 0 || pt.y < 0 ||
pt.x + _cascade->real_window_size.width >= hidCascade->sum->cols-2 ||
pt.y + _cascade->real_window_size.height >= hidCascade->sum->rows-2 )
return -1; //计算特征点在积分图中的偏移，相当于移动窗体
p_offset = pt.y * (hidCascade->sum->cols) + pt.x;
pq_offset = pt.y * (hidCascade->sqsum->cols) + pt.x; //计算移动后整个窗体的特征值
rectsum = calc_sum(*hidCascade,p_offset);//*cascade->inv_window_area;
variance_factor = hidCascade->pq0[pq_offset] - hidCascade->pq1[pq_offset] -
hidCascade->pq2[pq_offset] + hidCascade->pq3[pq_offset];
variance_factor = (variance_factor - ((rectsum*rectsum*windowArea[hidCascade->window_area-324])>>16))*windowArea[hidCascade->window_area-324]>>16;
//variance_norm_factor = int(sqrt(float(variance_factor))+0.5f);//qmath
variance_norm_factor = shortSqrtTable[variance_factor]; if( variance_norm_factor < 0 )
variance_norm_factor = 1; //计算每一个classifier的用到的特征区域的特征值 for( i = start_stage; i < hidCascade->count; i++ )
//for( i = start_stage; i < hidCascade->count; i++ )
{
stage_sum = 0; node = hidCascade->stage_classifier[i].classifier->node;
classifier = hidCascade->stage_classifier[i].classifier;
//if( hidCascade->stage_classifier[i].two_rects )
//{
for( j = 0; j < hidCascade->stage_classifier[i].count; j++ )
{
//start_time = _itoll(TSCH, TSCL);
//classifier = hidCascade->stage_classifier[i].classifier + j; //start_time = _itoll(TSCH, TSCL);
t = node->threshold*variance_norm_factor >> 10;
//end_time = _itoll(TSCH, TSCL);
// block1 += end_time - start_time - overhead; //start_time = _itoll(TSCH, TSCL);
//计算Haar特征
sum = calc_sum(node->feature.rect[0],p_offset) * node->feature.rect[0].weight >> 10;
sum += calc_sum(node->feature.rect[1],p_offset) * node->feature.rect[1].weight >> 10; //两特征和三特征分开处理
if( node->feature.rect[2].p0 )
sum += calc_sum(node->feature.rect[2],p_offset) * node->feature.rect[2].weight >> 10; //end_time = _itoll(TSCH, TSCL);
//block1 += end_time - start_time - overhead;
//
//a = classifier->alpha[0];
//b = classifier->alpha[1];
//start_time = _itoll(TSCH, TSCL);
stage_sum += sum < t ? classifier->alpha[0] : classifier->alpha[1];
// end_time = _itoll(TSCH, TSCL);
// block2 += end_time - start_time - overhead
node = (HidHaarTreeNode*)((char*)(node) + 80);
classifier++;
} if( stage_sum < hidCascade->stage_classifier[i].threshold )
{ return -i; }
} //QueryPerformanceCounter(&t2);
//printf("FeatureDetectTime:%fms\n",(t2.QuadPart - t1.QuadPart)*1000.0/tc.QuadPart); return 1;
} //==================================================================
//函数名： HaarDetectObjects
//作者： qiurenbo
//日期： 2014-09-30
//功能： 在指定图片中查找目标
//输入參数： _img 图片指针
// cascade 级联分类器指针
// start_stage 起始stage下标
// scale_factor32x 窗体变化尺度倍数 /32
// min_neighbors 最小临界目标（min_neighbors个以上的候选目标的区域才是最后的目标区域）
// minSize 目标最小的大小
//返回值： <=0 未检測到目标或參数有问题
// 1 成功检測到目标
//改动记录：
//====================================================================
void HaarDetectObjects(Image* _img,
HaarClassifierCascade* cascade, //训练好的级联分类器
char* storage, int scale_factor32x,
int min_neighbors, int flags, Size minSize)
{ //第一次分类用到的最大stage
//第二次分类用到的起始stage
int split_stage = 2; // ImgPtr stub, *img = _img;
Mat32 sum ;
Mat64 sqsum;
Image tmp; //检測区域候选队列
Sequence seq; //结果候选恿? Sequence seq2; //并查集合并序列
Sequence comps; Rect r1;
PTreeNode* node;
int r1_neighbor;
int j, flag = 1;
Rect r2 ;
int r2_neighbor;
int distance;//cvRound( r2.rect.width * 0.2 );
memset(&seq, 0, sizeof(Sequence));
memset(&comps, 0, sizeof(Sequence));
memset(&seq2, 0, sizeof(Sequence));
memset(&result_seq, 0, sizeof(result_seq)); int i; int factor32x;
int npass = 2; if( !cascade )
return ; //获取积分图和平方和积分图的矩阵
GetMat(&sum , _img->rows + 1, _img->cols + 1, BITS32, 0);
GetMat(&sqsum, _img->rows + 1, _img->cols + 1, BITS64, 0);
GetMat(&tmp, _img->rows, _img->cols, BITS8, 1); //若不存在隐式积分图（用于加速计算），则创建一个
if( !cascade->hid_cascade )
CreateHidHaarClassifierCascade(cascade); //计算积分图
Integral(_img, &sum, &sqsum); int count = 0;
int count2 = 0;
// In the variable declaration portion of the code:
/*uint64_t start_time, end_time, overhead, cyclecountSet=0, cyclecountRun=0;
// In the initialization portion of the code:
TSCL = 0; //enable TSC
start_time = _itoll(TSCH, TSCL);
end_time = _itoll(TSCH, TSCL);
overhead = end_time-start_time; //Calculating the overhead of the method.*/ //不断调整窗体尺度。直到到达图像边缘(_img->cols-10) ||(_img->rows - 10)
//而且确保尺度小于3倍(96)
for( factor32x = 32; factor32x*cascade->orig_window_size.width < (_img->cols - 10)<<5 &&
factor32x*cascade->orig_window_size.height < (_img->rows - 10)<<5
&&factor32x<96;
factor32x = (factor32x*scale_factor32x+16)>>5 )
{ const int ystep32x = MAX(64, factor32x); //调整搜索窗体尺度
Size win_size;
win_size.height = (cascade->orig_window_size.height * factor32x + 16)>>5;
win_size.width = (cascade->orig_window_size.width * factor32x + 16 )>>5; //pass指扫描次数，stage_offset指第二次扫描时从第几个stage開始
int pass, stage_offset = 0; //确保搜索窗体在尺度放大后仍然在图像中
int stop_height = ( ((_img->rows - win_size.height)<<5)+ (ystep32x>>1) ) / ystep32x; //确保搜索窗体大于目标的最小尺寸
if( win_size.width < minSize.width || win_size.height < minSize.height )
continue;
//QueryPerformanceFrequency(&tc);
//QueryPerformanceCounter(&t1);
//依据尺度设置隐式级联分类器中的特征和权重，并设置这些特征在积分图中的位置，以加速运算 // Code to be profiled
//start_time = _itoll(TSCH, TSCL);
SetImagesForHaarClassifierCascade(cascade, &sum, &sqsum, factor32x );
//end_time = _itoll(TSCH, TSCL);
//cyclecountSet = end_time-start_time-overhead;
//QueryPerformanceCounter(&t2);
//printf("SetImageFeatureRunTime:%fms\n",(t2.QuadPart - t1.QuadPart)*1000.0/tc.QuadPart); //设置粗检測所使用的起始分类器
cascade->hid_cascade->count = split_stage; //用检測窗体扫描两遍图像:
//第一遍通过级联两个stage粗略定位目标大致区域，对候选区域进行标定（利用tmp矩阵）
//第二遍对标定的候选区域进行完整筛选，将候选区域放置到队列中
for( pass = 0; pass < npass; pass++ )
{ for( int _iy = 0; _iy < stop_height; _iy++ )
{
//检測窗体纵坐标步长为2。保持不变
int iy = (_iy*ystep32x+16)>>5;
int _ix, _xstep = 1; //stop_width是指_ix迭代的上限，_ix还要*ystep32x才是真正的窗体坐标
int stop_width =( ((_img->cols - win_size.width)<<5) +ystep32x/2) / ystep32x;
unsigned char* mask_row = tmp.imgPtr + tmp.cols* iy; for( _ix = 0; _ix < stop_width; _ix += _xstep )
{ //检測窗体横坐标按步长为4開始移动，若没有检測到目标，则改变下一次步长为2
int ix = (_ix*ystep32x+16)>>5; // it really should be ystep //当前检測窗体左上角坐标
Point pt;
pt.x = ix;
pt.y = iy; //粗略检測
if( pass == 0 )
{ int result = 0;
_xstep = 2; //start_time = _itoll(TSCH, TSCL);
result = RunHaarClassifierCascade( cascade, pt, 0 );
//end_time = _itoll(TSCH, TSCL);
//cyclecountRun += end_time-start_time-overhead;
if( result > 0 )
{
if( pass < npass - 1 )
mask_row[ix] = 1; }
//没有检測到改变步长为2（看ix的值）
if( result < 0 )
_xstep = 1;
}
//第二次检測先前粗定位的坐标
else if( mask_row[ix] )
{
//start_time = _itoll(TSCH, TSCL);
int result = RunHaarClassifierCascade(cascade, pt, stage_offset);
// end_time = _itoll(TSCH, TSCL);
// cyclecountRun += end_time-start_time-overhead; //count2++;
//int result = 0;
if( result > 0 )
{
seq.rectQueue[seq.tail].height = win_size.height;
seq.rectQueue[seq.tail].width = win_size.width;
seq.rectQueue[seq.tail].x = ix;
seq.rectQueue[seq.tail].y = iy;
seq.total++;
seq.tail++;
}
else
mask_row[ix] = 0; }
} } //由于前两个stage在第一次检測的时候已经用过。
//第二次检測的时候。从第3个stage開始进行完整的检測
stage_offset = cascade->hid_cascade->count;
cascade->hid_cascade->count = cascade->count;
//cascade->hid_cascade->count = 15;
}
} //printf("The SetImage section took: %lld CPU cycles\n", cyclecountSet);
// printf("The RunImage section took: %lld CPU cycles\n", cyclecountRun);
// printf("The Block1 section took: %lld CPU cycles\n", block1);
// printf("The Block2 section took: %lld CPU cycles\n", block2); if( min_neighbors != 0 )
{ //将候选目标按类似度构成并查集
//返回值代表并查集树的个数
int ncomp = SeqPartition(&seq); //对相邻候选区域进行累加，为计算平均边界做准备
for( i = 0; i < seq.total; i++ )
{
r1 = seq.rectQueue[i];
node = &PTreeNodes[i];
while(node->parent)
node = node->parent;
int idx = (node - PTreeNodes); comps.neighbors[idx]++; comps.rectQueue[idx].x += r1.x;
comps.rectQueue[idx].y += r1.y;
comps.rectQueue[idx].width += r1.width;
comps.rectQueue[idx].height += r1.height;
} // 计算平均目标边界
for( i = 0; i < seq.total; i++ )
{
int n = comps.neighbors[i]; //仅仅有满足最小临接的结果才是终于结果
if( n >= min_neighbors )
{
Rect* rect = &seq2.rectQueue[seq2.tail];
rect->x = (comps.rectQueue[i].x*2 + n)/(2*n);
rect->y = (comps.rectQueue[i].y*2 + n)/(2*n);
rect->width = (comps.rectQueue[i].width*2 + n)/(2*n);
rect->height = (comps.rectQueue[i].height*2 + n)/(2*n);
seq2.neighbors[seq2.tail] = comps.neighbors[i];
seq2.tail++;
seq2.total++;
}
} //从候选矩形中得到最大的矩形
for( i = 0; i < seq2.total; i++ )
{
r1 = seq2.rectQueue[i];
r1_neighbor = seq2.neighbors[i];
flag = 1; for( j = 0; j < seq2.total; j++ )
{
r2 = seq2.rectQueue[j];
r2_neighbor = seq2.neighbors[j];
distance = (r2.width *2+5)/10;//cvRound( r2.rect.width * 0.2 ); if( i != j &&
r1.x >= r2.x - distance &&
r1.y >= r2.y - distance &&
r1.x + r1.width <= r2.x + r2.width + distance &&
r1.y + r1.height <= r2.y + r2.height + distance &&
(r2_neighbor > MAX( 3, r1_neighbor ) || r1_neighbor < 3) )
{
flag = 0;
break;
}
} if( flag )
{
result_seq.rectQueue[result_seq.tail] = r1;
result_seq.tail++;
result_seq.total++; }
} } } void DownSample(Image* pImage, int factor)
{
int i = 0;
int j = 0;
int counti = 0;
int countj = 0; int step = pImage->cols / factor;
for (i =0; i < pImage->rows; i+= factor)
{
countj++;
for (j =0; j < pImage->cols; j += factor)
{
*(pImage->imgPtr + i*step/factor + j/factor) = *(pImage->imgPtr + i*pImage->cols + j);
counti++;
}
counti = 0;
} pImage->cols /= factor;
pImage->rows /= factor;
}
Haar.h
#ifndef _HAAR_H_
#define _HAAR_H_
#include "Tables.h" #define NODE_THRESHOLD_SHIFT 22 #define MAXHIDCASCADE 200000 //隐式级联分类器所占空间(字节)
#define MAXROWS 400
#define MAXCOLS 400
#define MAXSTAGES 22
#define MAXCLASSIFER 213
#define MAXTREENODE 2
#define MAXALPHA 2
#define MAXSEQS 25
#define MaxMatNum 2
#define RGBCHANNEL 3
#define BITS8 0x00000001
#define BITS32 0x00000010
#define BITS64 0x00000100 #define CV_8TO16U_SQR(x) my8x16uSqrTab[(x)+128]
#define CLR_RESULT_QUEUE() result_seq.tail = 0;\
result_seq.total = 0; typedef unsigned char BYTE; typedef long long _int64;
typedef unsigned char (*ImgPtr);
typedef unsigned char (*Mat8Ptr);
typedef int (*Mat32Ptr);
typedef _int64 (*Mat64Ptr); /*****************并查集数据结构*******************************/
#define MAXPTREENODES 100
typedef struct PTreeNode
{
struct PTreeNode* parent;
char* element;
int rank;
}PTreeNode; /************************积分图变量***************************/
typedef int sumtype;
typedef _int64 sqsumtype; /************************************************************/
typedef struct Rect
{
int x;
int y;
int width;
int height;
}Rect; typedef struct
{
int width;
int height; }Size; typedef struct Image
{
ImgPtr imgPtr;
int rows;
int cols;
}Image; typedef struct Mat8
{
Mat8Ptr mat8Ptr;
int rows;
int cols;
}Mat8;
typedef struct Mat32
{
Mat32Ptr mat32Ptr;
int rows;
int cols;
}Mat32; typedef struct Mat64
{
Mat64Ptr mat64Ptr;
int rows;
int cols;
}Mat64; typedef struct Sequence
{
int total;
Rect rectQueue[MAXSEQS];
int neighbors[MAXSEQS];
int tail;
}Sequence; //Haar特征的数量
#define CV_HAAR_FEATURE_MAX 3 /*************HidHaar to Caculation Feature***********************************/
typedef struct HidHaarFeature
{
struct
{
sumtype *p0, *p1, *p2, *p3;
int weight;
}
rect[CV_HAAR_FEATURE_MAX];
}HidHaarFeature; typedef struct HidHaarTreeNode
{
HidHaarFeature feature;
int threshold;
int left;
int right;
}HidHaarTreeNode; typedef struct HidHaarClassifier
{
int count;
//CvHaarFeature* orig_feature; HidHaarTreeNode* node;
int* alpha;
//HidHaarTreeNode node[MAXTREENODE];
//int alpha[MAXALPHA];
}HidHaarClassifier; typedef struct HidHaarStageClassifier
{
int count;
int threshold;
HidHaarClassifier* classifier;
//HidHaarClassifier classifier[MAXCLASSIFER];
int two_rects; struct HidHaarStageClassifier* next;
struct HidHaarStageClassifier* child;
struct HidHaarStageClassifier* parent;
}HidHaarStageClassifier; typedef struct HidHaarClassifierCascade
{
int count;
int is_stump_based;
int has_tilted_features;
int is_tree;
int window_area;
Mat32* sum;
Mat64* sqsum;
HidHaarStageClassifier* stage_classifier;
//HidHaarStageClassifier stage_classifier[MAXSTAGES];
sqsumtype *pq0, *pq1, *pq2, *pq3;
sumtype *p0, *p1, *p2, *p3; void** ipp_stages;
}HidHaarClassifierCascade; /******************Haar Cascade*****************************************/
typedef struct HaarFeature
{
int tilted;
struct
{
Rect r;
int weight;
} rect[CV_HAAR_FEATURE_MAX];
}HaarFeature; typedef struct HaarClassifier
{
int count;
HaarFeature* haar_feature;
int* threshold;
int* left;
int* right;
int* alpha;
}HaarClassifier; typedef struct HaarStageClassifier
{
int count;
int threshold;
HaarClassifier* classifier; int next;
int child;
int parent;
}HaarStageClassifier; typedef struct HaarClassifierCascade
{
int flags;
int count;
Size orig_window_size;
Size real_window_size;
int scale32x;
HaarStageClassifier* stage_classifier;
HidHaarClassifierCascade* hid_cascade;
}HaarClassifierCascade; typedef struct CvAvgComp
{
Rect rect;
int neighbors;
}
CvAvgComp; typedef struct Point
{
int x;
int y;
}Point; /******************全局变量****************************************/
//cascade
extern HaarClassifierCascade *cascade ;
//extern HidHaarClassifierCascade hid_cascade; //32bits cell Mat
extern int MatPool32[MaxMatNum][MAXROWS][MAXCOLS];
//8bits cell
extern unsigned char MatPool8[MaxMatNum][MAXROWS][MAXCOLS]; //8bits*3 cell
extern unsigned char ImgRGBPool8[MaxMatNum][RGBCHANNEL][MAXROWS][MAXCOLS];
//64bits float cell
extern _int64 MatPool64[MaxMatNum][MAXROWS][MAXCOLS]; //分类器检測结果区域序列
extern Sequence result_seq; /********************全局函数******************************************/
extern void ReadFaceCascade();
extern void HaarDetectObjects(Image* _img,HaarClassifierCascade* cascade,
char* storage, int scale_factor32x,
int min_neighbors, int flags, Size minSize);
#endif
OpenCV Haar AdaBoost源代码改进（比EMCV快6倍）的更多相关文章
- OpenCV Haar AdaBoost源码改进据说是比EMCV快6倍
<pre name="code" class="cpp">#include "Haar.h" #include "lo ...
- 【图像处理】Haar Adaboost 检测自定义目标（视频车辆检测算法代码）
阅读须知 本博客涉及到的资源: 正样本:http://download.csdn.net/detail/zhuangxiaobin/7326197 负样本:http://download.csdn.n ...
- 6-9 Haar+adaboost人脸识别
我们重点分析了Haar特征的概念以及如何计算Haar特征,并介绍了Haar+Adaboost分类器它们的组合以及Adaboost分类器如何使用和训练.这节课我们将通过代码来实现一下Haar+Adabo ...
- Protobuf有没有比JSON快5倍？用代码来击破pb性能神话
转 http://www.sohu.com/a/136487507_505779 2017-04-26 07:58 程序设计 /58 /技术 导读:Google 的 Protocol Buffers ...
- grep之字符串搜索算法Boyer-Moore由浅入深（比KMP快3-5倍）
这篇长文历时近两天终于完成了,前两天帮网站翻译一篇文章“为什么GNU grep如此之快?”,里面提及到grep速度快的一个重要原因是使用了Boyer-Moore算法作为字符串搜索算法,兴趣之下就想了解 ...
- Hadoop3.0新特性介绍，比Spark快10倍的Hadoop3.0新特性
Hadoop3.0新特性介绍,比Spark快10倍的Hadoop3.0新特性 Apache hadoop 项目组最新消息,hadoop3.x以后将会调整方案架构,将Mapreduce 基于内存+io+ ...
- grep之字符串搜索算法Boyer-Moore由浅入深（比KMP快3-5倍）（转）
这篇长文历时近两天终于完成了,前两天帮网站翻译一篇文章“为什么GNU grep如此之快?”,里面提及到grep速度快的一个重要原因是使用了Boyer-Moore算法作为字符串搜索算法,兴趣之下就想了解 ...
- MySQL 8.0 正式版 8.0.11 发布：比 MySQL 5.7 快 2 倍
ySQL 8.0 正式版 8.0.11 已发布,官方表示 MySQL 8 要比 MySQL 5.7 快 2 倍,还带来了大量的改进和更快的性能! 注意:从 MySQL 5.7 升级到 MySQL 8. ...
- 斯坦福新深度学习系统 NoScope：视频对象检测快1000倍
以作备份,来源http://jiasuhui.com/archives/178954 本文由“新智元”(微信ID:AI_era)编译,来源:dawn.cs.stanford.edu,编译:刘小芹 斯坦 ...
随机推荐
- 改动已有gpg密钥的用户标识及凝视
/********************************************************************* * Author : Samson * Date ...
- Floodlight 处理交换机增加/移除过程
Floodlight 使用的是Netty架构,在Controller.java 入口函数中显示创建ServerBootstrap,设置套接字选项,ChannelPipeline,此时监听套接 ...
- 用Python实现QQ找茬游戏外挂工具
源地址:http://cpiz.net/blog/2012/03/a_qq_zhaocha_assistant_by_python/ (原创作品,转载请注明出处)好久没写技术相关的博文,这次写篇有意思 ...
- 14.3.2.3 Consistent Nonlocking Reads 一致性非锁定读
14.3.2.3 Consistent Nonlocking Reads 一致性非锁定读 一致性读 意味着 InnoDB 使用多版本来保护查询一个数据库在当前时间点的快照. 查询看到被事务做出的修改, ...
- Windbg抓取程序崩溃的dmp文件的方法
Windbg抓取程序崩溃的dmp文件的方法 一. 简介 windbg是在windows平台下,强大的用户态和内核态调试工具.相比较于Visual Studio,它是一个轻量级的调试工具, ...
- Python Object Graphs — objgraph 1.7.2 documentation
Python Object Graphs - objgraph 1.7.2 documentation Python Object Graphs¶ objgraph is a module that ...
- Button UI Kit CSS3美丽Buttonbutton
<!DOCTYPE HTML> <html lang="en-US"> <head> <meta charset="UTF-8& ...
- 深度学习系列之CNN核心内容
导读 怎么样来理解近期异常火热的深度学习网络?深度学习有什么亮点呢?答案事实上非常简答.今年十月份有幸參加了深圳高交会的中科院院士论坛.IEEE fellow汤晓欧做了一场精彩的报告,这个问题被汤大神 ...
- Caused by: org.springframework.beans.factory.NoSuchBeanDefinitionException
1.错误描写叙述 usage: java org.apache.catalina.startup.Catalina [ -config {pathname} ] [ -nonaming ] { -he ...
- [Android学习笔记]PopupWindow的使用
什么时候使用PopupWindow? 当业务需求的交互形式需要在当前页弹出一个简单可选项UI与用户进行交互时,可使用PopupWindow完成此功能开发 Android Dev API Doc Pop ...