train _cascade 源码阅读之HOG特征

本文讨论OpenCV  train_cascade 级联分类器中的HOG特征实现，HOG特征原理可以参考此文。特征的初始化框架和LBP 特征是一致的，感兴趣可以参考

train_cascade 源码阅读之LBP 特征 中的介绍。
HOG，即Histogram of Oriented Gradient 方向梯度直方图，常用于解决人体目标的检测的图像描述子，用来表达人体，提取人体外形信息和运动信息形成丰富的特征集。
生成过程： 检测窗口--> 归一化图像--> 计算梯度-->对每一个cell块对梯度直方图进行规定权重的投影 --> 对每个重叠block块内的cell进行对比度归一化 --> 把所有block内的直方图向量一起组合成一个大的HOG特征向量。（参考自 blog.sina.com.cn/s/blog_60e6e3d50101bkpn.html）
在HOG特征的操作中，与前述的Haar特征与LBP特征不同的是，初始时矩阵不再是单一的一个sum，而是一个矩阵向量hist，其中含有9个类似sum一样的矩阵，分别存放每一个方向的结果。

class CvHOGEvaluator : public CvFeatureEvaluator
{
public:
……
    virtual void setImage(const cv::Mat& img, uchar clsLabel, int idx);
    virtual float operator()(
            int varIdx, 
            int sampleIdx) const;
……
protected:
    virtual void generateFeatures();
    virtual void integralHistogram(
            const cv::Mat &img, 
            std::vector<cv::Mat> &histogram, 
            cv::Mat &norm, 
            int nbins) const;
    class Feature
    {
    public:
        Feature();
        Feature( int offset, int x, int y, int cellW, int cellH );
        float calc( 
                const std::vector<cv::Mat> &_hists, 
                const cv::Mat &_normSum, size_t y, 
                int featComponent ) const;
        void write( cv::FileStorage &fs ) const;
        void write( cv::FileStorage &fs, int varIdx ) const;

        cv::Rect rect[N_CELLS]; //cells

        struct
        {
            int p0, p1, p2, p3;
        } fastRect[N_CELLS];
    };
    std::vector<Feature> features;

    cv::Mat normSum; //for nomalization calculation (L1 or L2)
    std::vector<cv::Mat> hist;
};

接下来是初始化积分图操作下标的过程。

void CvHOGEvaluator::generateFeatures()
{
    int offset = winSize.width + 1;
    Size blockStep;
    int x, y, t, w, h;

    for (t = 8; t <= winSize.width/2; t+=8) 
        //t = size of a cell. blocksize = 4*cellSize
    {
        blockStep = Size(4,4);
        w = 2*t; //width of a block
        h = 2*t; //height of a block
        for (x = 0; x <= winSize.width - w; x += blockStep.width)
        {
            for (y = 0; y <= winSize.height - h; y += blockStep.height)
            {
                features.push_back(Feature(offset, x, y, t, t));
            }
        }
        w = 2*t;
        h = 4*t;
        for (x = 0; x <= winSize.width - w; x += blockStep.width)
        {
            for (y = 0; y <= winSize.height - h; y += blockStep.height)
            {
                features.push_back(Feature(offset, x, y, t, 2*t));
            }
        }
        w = 4*t;
        h = 2*t;
        for (x = 0; x <= winSize.width - w; x += blockStep.width)
        {
            for (y = 0; y <= winSize.height - h; y += blockStep.height)
            {
                features.push_back(Feature(offset, x, y, 2*t, t));
            }
        }
    }

    numFeatures = (int)features.size();
}

t表示cell的尺寸，一个block含有2×2个cell，因此，t需要不大于winSize.width/2，在这里采用了三种不同的形状，block遍历的step是4×4的。输入给Feature的构造参数是偏移量，左上角坐标点，和cell的宽高。
接下来看Feature的构造。

CvHOGEvaluator::Feature::Feature( int offset, int x, int y, int cellW, int cellH )
{
    rect[0] = Rect(x, y, cellW, cellH); //cell0
    rect[1] = Rect(x+cellW, y, cellW, cellH); //cell1
    rect[2] = Rect(x, y+cellH, cellW, cellH); //cell2
    rect[3] = Rect(x+cellW, y+cellH, cellW, cellH); //cell3

    for (int i = 0; i < N_CELLS; i++)
    {
        CV_SUM_OFFSETS(fastRect[i].p0, fastRect[i].p1, fastRect[i].p2, fastRect[i].p3, rect[i], offset);
    }
}

分别创建了四个cell矩形，CV_SUM_OFFSET宏计算的是矩形上的点在拉成行向量的积分图中的偏移量。
积分图中的坐标算好了，再看积分图的生成过程。

void CvHOGEvaluator::
setImage(const Mat &img, uchar clsLabel, int idx)
{
    CV_DbgAssert( !hist.empty());
    CvFeatureEvaluator::setImage( img, clsLabel, idx );
    vector<Mat> integralHist;
    for (int bin = 0; bin < N_BINS; bin++)
    {
        integralHist.push_back( 
                    Mat(winSize.height + 1, 
                        winSize.width + 1, 
                        hist[bin].type(), 
                        hist[bin].ptr<float>((int)idx)) );
    }
    Mat integralNorm(
                winSize.height + 1, 
                winSize.width + 1, 
                normSum.type(), 
                normSum.ptr<float>((int)idx));
    integralHistogram(img, integralHist, integralNorm, (int)N_BINS);
}

与LBP，Haar相同，新建Mat，传入积分图的数据地址，不同的是这里是矩阵向量，保存9个方向的积分图。这里没有使用OpenCV自带的integral直接计算积分图，而是自行实现了一个。

void CvHOGEvaluator::integralHistogram(
        const Mat   &img, 
        vector<Mat> &histogram, 
        Mat         &norm, 
        int         nbins) const
{
    CV_Assert( img.type() == CV_8U || img.type() == CV_8UC3 );
    int x, y, binIdx;

    Size gradSize(img.size());
    Size histSize(histogram[0].size());
    Mat grad(gradSize, CV_32F);
    Mat qangle(gradSize, CV_8U);

    AutoBuffer<int> mapbuf(gradSize.width + gradSize.height + 4);
    int* xmap = (int*)mapbuf + 1;
    int* ymap = xmap + gradSize.width + 2;

    const int borderType = (int)BORDER_REPLICATE;

    for( x = -1; x < gradSize.width + 1; x++ )
        xmap[x] = borderInterpolate(x, gradSize.width, borderType);
    for( y = -1; y < gradSize.height + 1; y++ )
        ymap[y] = borderInterpolate(y, gradSize.height, borderType);

    int width = gradSize.width;
    AutoBuffer<float> _dbuf(width*4);
    float* dbuf = _dbuf;
    Mat Dx(1, width, CV_32F, dbuf);
    Mat Dy(1, width, CV_32F, dbuf + width);
    Mat Mag(1, width, CV_32F, dbuf + width*2);
    Mat Angle(1, width, CV_32F, dbuf + width*3);

    float angleScale = (float)(nbins/CV_PI);

    for( y = 0; y < gradSize.height; y++ )
    {
        const uchar* currPtr = img.data + img.step*ymap[y];
        const uchar* prevPtr = img.data + img.step*ymap[y-1];
        const uchar* nextPtr = img.data + img.step*ymap[y+1];
        float* gradPtr = (float*)grad.ptr(y);
        uchar* qanglePtr = (uchar*)qangle.ptr(y);

        for( x = 0; x < width; x++ )
        {
            dbuf[x] = (float)(currPtr[xmap[x+1]] - currPtr[xmap[x-1]]);
            dbuf[width + x] = (float)(nextPtr[xmap[x]] - prevPtr[xmap[x]]);
        }
        cartToPolar( Dx, Dy, Mag, Angle, false );
        for( x = 0; x < width; x++ )
        {
            float mag = dbuf[x+width*2];
            float angle = dbuf[x+width*3];
            angle = angle*angleScale - 0.5f;
            int bidx = cvFloor(angle);
            angle -= bidx;
            if( bidx < 0 )
                bidx += nbins;
            else if( bidx >= nbins )
                bidx -= nbins;

            qanglePtr[x] = (uchar)bidx;
            gradPtr[x] = mag;
        }
    }
    integral(grad, norm, grad.depth());

    float* histBuf;
    const float* magBuf;
    const uchar* binsBuf;

    int binsStep = (int)( qangle.step / sizeof(uchar) );
    int histStep = (int)( histogram[0].step / sizeof(float) );
    int magStep = (int)( grad.step / sizeof(float) );
    for( binIdx = 0; binIdx < nbins; binIdx++ )
    {
        histBuf = (float*)histogram[binIdx].data;
        magBuf = (const float*)grad.data;
        binsBuf = (const uchar*)qangle.data;

        memset( histBuf, 0, histSize.width * sizeof(histBuf[0]) );
        histBuf += histStep + 1;
        for( y = 0; y < qangle.rows; y++ )
        {
            histBuf[-1] = 0.f;
            float strSum = 0.f;
            for( x = 0; x < qangle.cols; x++ )
            {
                if( binsBuf[x] == binIdx )
                    strSum += magBuf[x];
                histBuf[x] = histBuf[-histStep + x] + strSum;
            }
            histBuf += histStep;
            binsBuf += binsStep;
            magBuf += magStep;
        }
    }
}

看完了代码才知道，以前对HOG特征的理解是有偏差的，尤其是在梯度的计算上，犯了严重的想当然的错误。特征计算完成后，调用integral计算平方积分图，再根据角度，将幅值放到每个积分直方图中。
最后通过如下方式调用计算HOG特征，并进行归一化。

inline float CvHOGEvaluator::Feature::calc( 
        const std::vector<cv::Mat>& _hists, 
        const cv::Mat& _normSum, size_t y, 
        int featComponent ) const
{
    float normFactor;
    float res;

    int binIdx = featComponent % N_BINS;
    int cellIdx = featComponent / N_BINS;

    const float *phist = _hists[binIdx].ptr<float>((int)y);
    res = phist[fastRect[cellIdx].p0] 
            - phist[fastRect[cellIdx].p1] 
            - phist[fastRect[cellIdx].p2] 
            + phist[fastRect[cellIdx].p3];

    const float *pnormSum = _normSum.ptr<float>((int)y);
    normFactor = (float)(pnormSum[fastRect[0].p0] 
            - pnormSum[fastRect[1].p1] 
            - pnormSum[fastRect[2].p2] 
            + pnormSum[fastRect[3].p3]);
    res = (res > 0.001f) ? ( res / (normFactor + 0.001f) ) : 0.f; 
    //for cutting negative values, which apper due to floating precision

    return res;
}