浅谈OPENCV人脸识别(转)
OpenCV的人脸检测主要是调用训练好的cascade（Haar分类器）来进行模式匹配。
cvHaarDetectObjects，先将图像灰度化，根据传入参数判断是否进行canny边缘处理(默认不使用)，再进行匹配。匹配后收集找出的匹配块,过滤噪声，计算相邻个数如果超过了规定值（传入的min_neighbors）就当成输出结果，否则删去。
匹配循环：将匹配分类器放大scale（传入值）倍，同时原图缩小scale倍，进行匹配，直到匹配分类器的大小大于原图，则返回匹配结果。匹配的时候调用cvRunHaarClassifierCascade来进行匹配，将所有结果存入CvSeq* Seq （可动态增长元素序列），将结果传给cvHaarDetectObjects。
cvRunHaarClassifierCascade函数整体是根据传入的图像和cascade来进行匹配。并且可以根据传入的cascade类型不同（树型、stump（不完整的树）或其他的），进行不同的匹配方式。
函数 cvRunHaarClassifierCascade 用于对单幅图片的检测。在函数调用前首先利用 cvSetImagesForHaarClassifierCascade设定积分图和合适的比例系数 (=> 窗口尺寸)。当分析的矩形框全部通过级联分类器每一层的时返回正值(这是一个候选目标)，否则返回0或负值。
为了了解OpenCV人脸检测中寻找匹配图像的详细过程，就把cvHaarDetectObjects和cvRunHaarClassifierCascade的源文件详细看了一遍，并打上了注释。方便大家阅读。
附cvHaarDetectObjects代码：
CV_IMPL CvSeq*
cvHaarDetectObjects(const CvArr* _img,
CvHaarClassifierCascade* cascade,
CvMemStorage* storage, double scale_factor,
int min_neighbors, int flags, CvSize min_size )
{
int split_stage = 2;
CvMatstub, *img = (CvMat*)_img; //CvMat多通道矩阵 *img=_img指针代换传入图
CvMat *temp = 0, *sum = 0, *tilted = 0, *sqsum = 0, *norm_img = 0, *sumcanny =0, *img_small = 0;
CvSeq* seq = 0;
CvSeq*seq2 =0; //CvSeq可动态增长元素序列
CvSeq* idx_seq = 0;
CvSeq* result_seq = 0;
CvMemStorage* temp_storage = 0;
CvAvgComp* comps = 0;
int i;
#ifdef_OPENMP
CvSeq* seq_thread[CV_MAX_THREADS] = {0};
int max_threads = 0;
#endif
CV_FUNCNAME( “cvHaarDetectObjects” );
__BEGIN__;
double factor;
intnpass = 2,coi; //npass=2
int do_canny_pruning = flags &CV_HAAR_DO_CANNY_PRUNING; //true做canny边缘处理
if( !CV_IS_HAAR_CLASSIFIER(cascade) )
CV_ERROR( !cascade ? CV_StsNullPtr : CV_StsBadArg, “Invalid classifier cascade”);
if( !storage )
CV_ERROR( CV_StsNullPtr, “Null storage pointer” );
CV_CALL( img = cvGetMat( img, &stub, &coi ));
if( coi )
CV_ERROR( CV_BadCOI, “COI is not supported” ); //一些出错代码
if( CV_MAT_DEPTH(img->type) != CV_8U )
CV_ERROR( CV_StsUnsupportedFormat, “Only 8-bit images are supported” );
CV_CALL( temp = cvCreateMat( img->rows, img->cols, CV_8UC1 ));
CV_CALL( sum = cvCreateMat( img->rows + 1, img->cols + 1, CV_32SC1 ));
CV_CALL( sqsum = cvCreateMat( img->rows + 1, img->cols + 1, CV_64FC1 ));
CV_CALL( temp_storage = cvCreateChildMemStorage( storage ));
#ifdef _OPENMP
max_threads = cvGetNumThreads();
for( i = 0; i < max_threads; i++ )
{
CvMemStorage* temp_storage_thread;
CV_CALL( temp_storage_thread =cvCreateMemStorage(0)); //CV_CALL就是运行，假如出错就报错。
CV_CALL( seq_thread[i] = cvCreateSeq( 0,sizeof(CvSeq), //CvSeq可动态增长元素序列
sizeof(CvRect),temp_storage_thread ));
}
#endif
if( !cascade->hid_cascade )
CV_CALL( icvCreateHidHaarClassifierCascade(cascade) );
if( cascade->hid_cascade->has_tilted_features )
tilted = cvCreateMat( img->rows + 1, img->cols + 1, CV_32SC1); //多通道矩阵图像长宽+1 4通道
seq = cvCreateSeq( 0, sizeof(CvSeq), sizeof(CvRect), temp_storage); //创建序列seq 矩形
seq2 = cvCreateSeq( 0, sizeof(CvSeq), sizeof(CvAvgComp), temp_storage); //创建序列seq2 矩形和邻近
result_seq = cvCreateSeq( 0, sizeof(CvSeq), sizeof(CvAvgComp), storage); //创建序列result_seq 矩形和邻近
if( min_neighbors == 0 )
seq = result_seq;
if( CV_MAT_CN(img->type) > 1 )
{
cvCvtColor( img, temp, CV_BGR2GRAY ); //img转为灰度
img =temp;
}
if( flags & CV_HAAR_SCALE_IMAGE) //flag&& 匹配图
{
CvSize win_size0 = cascade->orig_window_size; //CvSize win_size0为分类器的原始大小
int use_ipp = cascade->hid_cascade->ipp_stages != 0&&
icvApplyHaarClassifier_32s32f_C1R_p!= 0; //IPP相关函数
if( use_ipp )
CV_CALL( norm_img = cvCreateMat( img->rows, img->cols, CV_32FC1)); //图像的矩阵化 4通道.
CV_CALL( img_small = cvCreateMat( img->rows + 1, img->cols + 1, CV_8UC1)); //小图矩阵化单通道长宽+1
for( factor = 1; ;factor *= scale_factor ) //成scale_factor倍数匹配
{
int positive = 0;
int x, y;
CvSize win_size = { cvRound(win_size0.width*factor),
cvRound(win_size0.height*factor) }; //winsize 分类器行列（扩大factor倍）
CvSize sz = { cvRound( img->cols/factor ), cvRound( img->rows/factor )}; //sz 图像行列（缩小factor倍）三个Cvsize
CvSize sz1 = { sz.width – win_size0.width, sz.height – win_size0.height}; //sz1 图像减分类器行列
CvRect rect1 = { icv_object_win_border, icv_object_win_border,
win_size0.width – icv_object_win_border*2, //icv_object_win_border（int）初始值=1
win_size0.height –icv_object_win_border*2 }; //矩形框rect1
CvMatimg1, sum1, sqsum1, norm1, tilted1, mask1; //多通道矩阵
CvMat* _tilted = 0;
if( sz1.width <= 0 || sz1.height <= 0 ) //图片宽或高小于分类器–>跳出
break;
if( win_size.width < min_size.width || win_size.height < min_size.height) //分类器高或宽小于给定的mini_size的高或宽–>继续
continue;
//CV_8UC1见定义.
//#defineCV_MAKETYPE(depth,cn) ((depth) + (((cn)-1) <<CV_CN_SHIFT))
//深度+（cn-1）左移3位 depth，depth+8，depth+16，depth+24.
img1 = cvMat( sz.height, sz.width, CV_8UC1, img_small->data.ptr); //小图的矩阵化 img1 单通道
sum1 = cvMat( sz.height+1, sz.width+1, CV_32SC1, sum->data.ptr); //长宽+1 4通道8位 多通道矩阵
sqsum1 = cvMat( sz.height+1, sz.width+1, CV_64FC1, sqsum->data.ptr); //长宽+1 4通道16位
if( tilted )
{
tilted1 = cvMat( sz.height+1, sz.width+1, CV_32SC1, tilted->data.ptr); //长宽+1 4通道8位
_tilted =&tilted1; //长宽+1 4通道8位
}
norm1 = cvMat( sz1.height, sz1.width, CV_32FC1, norm_img ?norm_img->data.ptr : 0 ); //norm1 图像减分类器行列 4通道
mask1 = cvMat( sz1.height, sz1.width, CV_8UC1, temp->data.ptr); //mask1 灰度图
cvResize( img, &img1, CV_INTER_LINEAR); //img双线性插值输出到img1
cvIntegral( &img1, &sum1, &sqsum1, _tilted); //计算积分图像
if( use_ipp && icvRectStdDev_32s32f_C1R_p( sum1.data.i, sum1.step,
sqsum1.data.db, sqsum1.step, norm1.data.fl, norm1.step, sz1, rect1 ) < 0 )
use_ipp = 0;
if( use_ipp) //如果ipp=true (intel视频处理加速等的函数库)
{
positive =mask1.cols*mask1.rows; //mask1长乘宽–>positive
cvSet( &mask1, cvScalarAll(255) ); //mask1赋值为255
for( i = 0; i < cascade->count; i++ )
{
if( icvApplyHaarClassifier_32s32f_C1R_p(sum1.data.i, sum1.step,
norm1.data.fl, norm1.step, mask1.data.ptr, mask1.step,
sz1, &positive, cascade->hid_cascade->stage_classifier[i].threshold,
cascade->hid_cascade->ipp_stages[i]) < 0 )
{
use_ipp=0; //ipp=false;
break;
}
if( positive <= 0 )
break;
}
}
if( !use_ipp) //如果ipp=false
{
cvSetImagesForHaarClassifierCascade( cascade, &sum1, &sqsum1, 0, 1. );
for( y = 0, positive = 0; y < sz1.height; y++ )
for( x = 0; x < sz1.width; x++ )
{
mask1.data.ptr[mask1.step*y + x] =
cvRunHaarClassifierCascade( cascade, cvPoint(x,y), 0 ) > 0; //匹配图像.
positive += mask1.data.ptr[mask1.step*y + x];
}
}
if( positive > 0 )
{
for( y = 0; y < sz1.height; y++ )
for( x = 0; x < sz1.width; x++ )
if( mask1.data.ptr[mask1.step*y + x] != 0 )
{
CvRect obj_rect = { cvRound(y*factor),cvRound(x*factor),
win_size.width,win_size.height };
cvSeqPush( seq, &obj_rect ); //将匹配块放到seq中
}
}
}
}
else //!(flag&& 匹配图)
{
cvIntegral( img, sum, sqsum, tilted );
if( do_canny_pruning )
{
sumcanny = cvCreateMat( img->rows + 1, img->cols + 1, CV_32SC1); //如果做canny边缘检测
cvCanny( img, temp, 0, 50, 3 );
cvIntegral( temp, sumcanny );
}
if( (unsigned)split_stage >= (unsigned)cascade->count ||
cascade->hid_cascade->is_tree) {
split_stage = cascade->count;
npass = 1;
}
for( factor = 1; factor*cascade->orig_window_size.width < img->cols –10&& //匹配
factor*cascade->orig_window_size.height < img->rows – 10;
factor *= scale_factor )
{
const double ystep = MAX( 2, factor );
CvSize win_size = { cvRound( cascade->orig_window_size.width * factor ),
cvRound( cascade->orig_window_size.height * factor )};
CvRect equ_rect = { 0, 0, 0, 0 };
int *p0 = 0, *p1 = 0, *p2 = 0, *p3 = 0;
int *pq0 = 0, *pq1 = 0, *pq2 = 0, *pq3 = 0;
int pass, stage_offset = 0;
int stop_height = cvRound((img->rows – win_size.height) / ystep);
if( win_size.width < min_size.width || win_size.height < min_size.height) //超边跳出
continue;
cvSetImagesForHaarClassifierCascade( cascade, sum, sqsum, tilted, factor); //匹配
cvZero( temp); //清空temp数组
if( do_canny_pruning) //canny边缘检测
{
equ_rect.x = cvRound(win_size.width*0.15);
equ_rect.y = cvRound(win_size.height*0.15);
equ_rect.width = cvRound(win_size.width*0.7);
equ_rect.height = cvRound(win_size.height*0.7);
p0 = (int*)(sumcanny->data.ptr + equ_rect.y*sumcanny->step) + equ_rect.x;
p1 = (int*)(sumcanny->data.ptr + equ_rect.y*sumcanny->step)
+ equ_rect.x + equ_rect.width;
p2 = (int*)(sumcanny->data.ptr + (equ_rect.y +equ_rect.height)*sumcanny->step) + equ_rect.x;
p3 = (int*)(sumcanny->data.ptr + (equ_rect.y +equ_rect.height)*sumcanny->step)
+ equ_rect.x + equ_rect.width;
pq0 = (int*)(sum->data.ptr + equ_rect.y*sum->step) + equ_rect.x;
pq1 = (int*)(sum->data.ptr + equ_rect.y*sum->step)
+ equ_rect.x + equ_rect.width;
pq2 = (int*)(sum->data.ptr + (equ_rect.y + equ_rect.height)*sum->step) +equ_rect.x;
pq3 = (int*)(sum->data.ptr + (equ_rect.y + equ_rect.height)*sum->step)
+ equ_rect.x + equ_rect.width;
}
cascade->hid_cascade->count =split_stage; //分裂级
for( pass = 0; pass < npass; pass++ )
{
#ifdef _OPENMP
#pragma omp parallel for num_threads(max_threads), schedule(dynamic)
#endif
for( int _iy = 0; _iy < stop_height; _iy++ )
{
int iy = cvRound(_iy*ystep);
int _ix, _xstep = 1;
int stop_width = cvRound((img->cols – win_size.width) / ystep);
uchar* mask_row = temp->data.ptr + temp->step * iy;
for( _ix = 0; _ix < stop_width; _ix += _xstep )
{
int ix = cvRound(_ix*ystep); // it really should be ystep
if( pass == 0 ) //第一次循环做
{
int result;
_xstep = 2;
if( do_canny_pruning ) //canny边缘检测
{
int offset;
int s, sq;
offset = iy*(sum->step/sizeof(p0[0])) + ix;
s= p0[offset] – p1[offset] – p2[offset] + p3[offset];
sq= pq0[offset] – pq1[offset] – pq2[offset] + pq3[offset];
if( s < 100 || sq < 20 )
continue;
}
result = cvRunHaarClassifierCascade( cascade, cvPoint(ix,iy), 0); //匹配结果存到result里
if( result > 0 )
{
if( pass < npass – 1 )
mask_row[ix] = 1;
else
{
CvRect rect = cvRect(ix,iy,win_size.width,win_size.height);
#ifndef_OPENMP //如果用OpenMP
cvSeqPush( seq, &rect ); //result放到seq中
#else //如果不用OpenMP
cvSeqPush( seq_thread[omp_get_thread_num()], &rect); //result放到seq_thread里
#endif
}
}
if( result < 0 )
_xstep = 1;
}
else if( mask_row[ix]) //不是第一次
{
int result = cvRunHaarClassifierCascade( cascade,cvPoint(ix,iy), stage_offset );
if( result > 0 )
{
if( pass == npass – 1 ) //如果是最后一次
{
CvRect rect = cvRect(ix,iy,win_size.width,win_size.height);
#ifndef _OPENMP
cvSeqPush( seq, &rect );
#else
cvSeqPush( seq_thread[omp_get_thread_num()], &rect );
#endif
}
}
else
mask_row[ix] = 0;
}
}
}
stage_offset = cascade->hid_cascade->count;
cascade->hid_cascade->count = cascade->count;
}
}
}
#ifdef _OPENMP
// gather theresults //收集结果
for( i = 0; i <max_threads; i++ )
{
CvSeq* s =seq_thread[i];
int j, total = s->total;
CvSeqBlock* b = s->first;
for( j = 0; j < total; j += b->count, b = b->next )
cvSeqPushMulti( seq, b->data, b->count); //结果输出到seq
}
#endif
if( min_neighbors != 0 )
{// group retrieved rectangles in order to filter outnoise 收集找出的匹配块,过滤噪声
int ncomp = cvSeqPartition( seq, 0, &idx_seq, is_equal, 0 );
CV_CALL( comps = (CvAvgComp*)cvAlloc( (ncomp+1)*sizeof(comps[0])));
memset( comps, 0, (ncomp+1)*sizeof(comps[0]));
// count number ofneighbors 计算相邻个数
for( i = 0; i < seq->total; i++ )
{
CvRect r1 = *(CvRect*)cvGetSeqElem( seq, i );
int idx = *(int*)cvGetSeqElem( idx_seq, i );
assert( (unsigned)idx < (unsigned)ncomp );
comps[idx].neighbors++;
comps[idx].rect.x += r1.x;
comps[idx].rect.y += r1.y;
comps[idx].rect.width += r1.width;
comps[idx].rect.height += r1.height;
}
// calculate average boundingbox 计算重心
for( i = 0; i < ncomp; i++ )
{
int n = comps[i].neighbors;
if( n >= min_neighbors )
{
CvAvgComp comp;
comp.rect.x = (comps[i].rect.x*2 + n)/(2*n);
comp.rect.y = (comps[i].rect.y*2 + n)/(2*n);
comp.rect.width = (comps[i].rect.width*2 + n)/(2*n);
comp.rect.height = (comps[i].rect.height*2 + n)/(2*n);
comp.neighbors = comps[i].neighbors;
cvSeqPush( seq2, &comp ); //结果输入到seq2
}
}
// filter out small face rectangles inside large face rectangles 在大的面块中找出小的面块
for( i = 0; i < seq2->total; i++) //在seq2中寻找
{
CvAvgComp r1 = *(CvAvgComp*)cvGetSeqElem( seq2, i); //r1指向结果
int j, flag = 1;
for( j = 0; j < seq2->total; j++ )
{
CvAvgComp r2 = *(CvAvgComp*)cvGetSeqElem( seq2, j );
int distance = cvRound( r2.rect.width * 0.2 );
if( i != j &&
r1.rect.x >= r2.rect.x – distance &&
r1.rect.y >= r2.rect.y – distance &&
r1.rect.x + r1.rect.width <= r2.rect.x + r2.rect.width + distance &&
r1.rect.y + r1.rect.height <= r2.rect.y + r2.rect.height + distance&&
(r2.neighbors > MAX( 3, r1.neighbors ) || r1.neighbors < 3) )
{
flag = 0;
break;
}
}
if( flag )
{
cvSeqPush( result_seq, &r1 ); //添加r1到返回结果.
/* cvSeqPush( result_seq, &r1.rect ); */
}
}
}
__END__;
#ifdef _OPENMP
for( i = 0; i <max_threads; i++ )
{
if( seq_thread[i] )
cvReleaseMemStorage( &seq_thread[i]->storage); //如果使用了OpenMP就释放使用的seq_thread
}
#endif
cvReleaseMemStorage( &temp_storage );
cvReleaseMat( &sum );
cvReleaseMat( &sqsum );
cvReleaseMat( &tilted ); //释放使用的空间
cvReleaseMat( &temp );
cvReleaseMat( &sumcanny );
cvReleaseMat( &norm_img );
cvReleaseMat( &img_small );
cvFree( &comps );
returnresult_seq; //返回结果
}
下面是cvRunHaarClassifierCascade的：
CV_IMPL int
cvRunHaarClassifierCascade( CvHaarClassifierCascade* _cascade,
CvPoint pt, int start_stage )
{
int result = -1;
CV_FUNCNAME(”cvRunHaarClassifierCascade”);
__BEGIN__;
int p_offset, pq_offset;
int i, j;
double mean, variance_norm_factor;
CvHidHaarClassifierCascade* cascade;
if( !CV_IS_HAAR_CLASSIFIER(_cascade) )
CV_ERROR( !_cascade ? CV_StsNullPtr: CV_StsBadArg, “Invalid cascade pointer” );
cascade = _cascade->hid_cascade;
if( !cascade )
CV_ERROR( CV_StsNullPtr, “Hiddencascade has not been created.\n”
“UsecvSetImagesForHaarClassifierCascade” );
if( pt.x < 0 || pt.y < 0 ||
pt.x +_cascade->real_window_size.width >= cascade->sum.width-2 ||
pt.y + _cascade->real_window_size.height>= cascade->sum.height-2) //超边退出
EXIT;
p_offset = pt.y * (cascade->sum.step/sizeof(sumtype)) + pt.x;
pq_offset = pt.y *(cascade->sqsum.step/sizeof(sqsumtype)) + pt.x;
mean = calc_sum(*cascade,p_offset)*cascade->inv_window_area;
variance_norm_factor = cascade->pq0[pq_offset] –cascade->pq1[pq_offset]- //左上+右下-右上-左下
cascade->pq2[pq_offset] + cascade->pq3[pq_offset];
variance_norm_factor = variance_norm_factor*cascade->inv_window_area– mean*mean;
if( variance_norm_factor >= 0. )
variance_norm_factor =sqrt(variance_norm_factor);
else
variance_norm_factor = 1.;
if( cascade->is_tree) //是树形的分类器,就按照层来匹配.
{
CvHidHaarStageClassifier* ptr;
assert( start_stage == 0); //start_stage==0继续
result = 1;
ptr = cascade->stage_classifier;
while( ptr )
{
double stage_sum =0;
for( j = 0; j < ptr->count; j++ )
{ stage_sum += icvEvalHidHaarClassifier( ptr->classifier +j, //层判断
variance_norm_factor, p_offset );
}
if( stage_sum >= ptr->threshold )
{
ptr =ptr->child; //层判断通过,到下一层.
}
else
{
while( ptr && ptr->next == NULL ) ptr =ptr->parent; //未通过,且当前子分类器没有同层分类器,没有返回上层
if( ptr == NULL) //如果刚才已经是最顶层了.
{
result =0; //返回0,退出.
EXIT;
}
ptr =ptr->next; //指向下一个分类器.
}
}
}
else if( cascade->is_stump_based) //如果是stump类的分类器
{
for( i = start_stage; i <cascade->count; i++ )
{
doublestage_sum = 0;
if( cascade->stage_classifier[i].two_rects )
{
for( j = 0; j < cascade->stage_classifier[i].count; j++ )
{
CvHidHaarClassifier* classifier = cascade->stage_classifier[i].classifier +j;
CvHidHaarTreeNode* node = classifier->node;
double sum, t = node->threshold*variance_norm_factor, a, b;
sum = calc_sum(node->feature.rect[0],p_offset) *node->feature.rect[0].weight;
sum += calc_sum(node->feature.rect[1],p_offset) *node->feature.rect[1].weight;
a = classifier->alpha[0];
b = classifier->alpha[1];
stage_sum += sum < t ? a : b;
}
}
else
{
for( j = 0; j < cascade->stage_classifier[i].count; j++ )
{
CvHidHaarClassifier* classifier = cascade->stage_classifier[i].classifier +j;
CvHidHaarTreeNode* node = classifier->node;
double sum, t = node->threshold*variance_norm_factor, a, b;
sum = calc_sum(node->feature.rect[0],p_offset) *node->feature.rect[0].weight;
sum += calc_sum(node->feature.rect[1],p_offset) *node->feature.rect[1].weight;
if( node->feature.rect[2].p0 )
sum += calc_sum(node->feature.rect[2],p_offset) *node->feature.rect[2].weight;
a = classifier->alpha[0];
b = classifier->alpha[1];
stage_sum += sum < t ? a : b;
}
}
if( stage_sum < cascade->stage_classifier[i].threshold )
{ //没通过.则返回负的没通过的分类器数.
result = -i;
EXIT;
}
}
}
else //如果不是那两种强分类器
{
for( i = start_stage; i <cascade->count; i++ )
{
double stage_sum =0;
for( j = 0; j < cascade->stage_classifier[i].count; j++ )
{
stage_sum +=icvEvalHidHaarClassifier(
cascade->stage_classifier[i].classifier + j,
variance_norm_factor, p_offset );
}
if( stage_sum < cascade->stage_classifier[i].threshold )
{
result = -i;
EXIT;
}
}
}
result = 1;
__END__;
return result; //返回结果
}
