【OpenCV】SIFT原理与源码分析:关键点搜索与定位

时间:2024-01-01 21:29:57

《SIFT原理与源码分析》系列文章索引:http://www.cnblogs.com/tianyalu/p/5467813.html

由前一步《DoG尺度空间构造》,我们得到了DoG高斯差分金字塔:

【OpenCV】SIFT原理与源码分析:关键点搜索与定位

如上图的金字塔,高斯尺度空间金字塔中每组有五层不同尺度图像,相邻两层相减得到四层DoG结果。关键点搜索就在这四层DoG图像上寻找局部极值点。

DoG局部极值点

寻找DoG极值点时,每一个像素点和它所有的相邻点比较,当其大于(或小于)它的图像域和尺度域的所有相邻点时,即为极值点。如下图所示,比较的范围是个3×3的立方体:中间的检测点和它同尺度的8个相邻点,以及和上下相邻尺度对应的9×2个点——共26个点比较,以确保在尺度空间和二维图像空间都检测到极值点。

【OpenCV】SIFT原理与源码分析:关键点搜索与定位

在一组中,搜索从每组的第二层开始,以第二层为当前层,第一层和第三层分别作为立方体的的上下层;搜索完成后再以第三层为当前层做同样的搜索。所以每层的点搜索两次。通常我们将组Octaves索引以-1开始,则在比较时牺牲了-1组的第0层和第N组的最高层

【OpenCV】SIFT原理与源码分析:关键点搜索与定位

高斯金字塔,DoG图像及极值计算的相互关系如上图所示。

关键点精确定位

以上极值点的搜索是在离散空间进行搜索的,由下图可以看到,在离散空间找到的极值点不一定是真正意义上的极值点。可以通过对尺度空间DoG函数进行曲线拟合寻找极值点来减小这种误差。
【OpenCV】SIFT原理与源码分析:关键点搜索与定位
利用DoG函数在尺度空间的Taylor展开式:
【OpenCV】SIFT原理与源码分析:关键点搜索与定位
则极值点为:
【OpenCV】SIFT原理与源码分析:关键点搜索与定位
程序中还除去了极值小于0.04的点。如下所示:
// Detects features at extrema in DoG scale space.  Bad features are discarded
// based on contrast and ratio of principal curvatures.
// 在DoG尺度空间寻特征点(极值点)
void SIFT::findScaleSpaceExtrema( const vector<Mat>& gauss_pyr, const vector<Mat>& dog_pyr,
vector<KeyPoint>& keypoints ) const
{
int nOctaves = (int)gauss_pyr.size()/(nOctaveLayers + ); // The contrast threshold used to filter out weak features in semi-uniform
// (low-contrast) regions. The larger the threshold, the less features are produced by the detector.
// 过滤掉弱特征的阈值 contrastThreshold默认为0.04
int threshold = cvFloor(0.5 * contrastThreshold / nOctaveLayers * * SIFT_FIXPT_SCALE);
const int n = SIFT_ORI_HIST_BINS; //
float hist[n];
KeyPoint kpt; keypoints.clear(); for( int o = ; o < nOctaves; o++ )
for( int i = ; i <= nOctaveLayers; i++ )
{
int idx = o*(nOctaveLayers+)+i;
const Mat& img = dog_pyr[idx];
const Mat& prev = dog_pyr[idx-];
const Mat& next = dog_pyr[idx+];
int step = (int)img.step1();
int rows = img.rows, cols = img.cols; for( int r = SIFT_IMG_BORDER; r < rows-SIFT_IMG_BORDER; r++)
{
const short* currptr = img.ptr<short>(r);
const short* prevptr = prev.ptr<short>(r);
const short* nextptr = next.ptr<short>(r); for( int c = SIFT_IMG_BORDER; c < cols-SIFT_IMG_BORDER; c++)
{
int val = currptr[c]; // find local extrema with pixel accuracy
// 寻找局部极值点,DoG中每个点与其所在的立方体周围的26个点比较
// if (val比所有都大 或者 val比所有都小)
if( std::abs(val) > threshold &&
((val > && val >= currptr[c-] && val >= currptr[c+] &&
val >= currptr[c-step-] && val >= currptr[c-step] &&
val >= currptr[c-step+] && val >= currptr[c+step-] &&
val >= currptr[c+step] && val >= currptr[c+step+] &&
val >= nextptr[c] && val >= nextptr[c-] &&
val >= nextptr[c+] && val >= nextptr[c-step-] &&
val >= nextptr[c-step] && val >= nextptr[c-step+] &&
val >= nextptr[c+step-] && val >= nextptr[c+step] &&
val >= nextptr[c+step+] && val >= prevptr[c] &&
val >= prevptr[c-] && val >= prevptr[c+] &&
val >= prevptr[c-step-] && val >= prevptr[c-step] &&
val >= prevptr[c-step+] && val >= prevptr[c+step-] &&
val >= prevptr[c+step] && val >= prevptr[c+step+]) ||
(val < && val <= currptr[c-] && val <= currptr[c+] &&
val <= currptr[c-step-] && val <= currptr[c-step] &&
val <= currptr[c-step+] && val <= currptr[c+step-] &&
val <= currptr[c+step] && val <= currptr[c+step+] &&
val <= nextptr[c] && val <= nextptr[c-] &&
val <= nextptr[c+] && val <= nextptr[c-step-] &&
val <= nextptr[c-step] && val <= nextptr[c-step+] &&
val <= nextptr[c+step-] && val <= nextptr[c+step] &&
val <= nextptr[c+step+] && val <= prevptr[c] &&
val <= prevptr[c-] && val <= prevptr[c+] &&
val <= prevptr[c-step-] && val <= prevptr[c-step] &&
val <= prevptr[c-step+] && val <= prevptr[c+step-] &&
val <= prevptr[c+step] && val <= prevptr[c+step+])))
{
int r1 = r, c1 = c, layer = i; // 关键点精确定位
if( !adjustLocalExtrema(dog_pyr, kpt, o, layer, r1, c1,
nOctaveLayers, (float)contrastThreshold,
(float)edgeThreshold, (float)sigma) )
continue; float scl_octv = kpt.size*0.5f/( << o);
// 计算梯度直方图
float omax = calcOrientationHist(
gauss_pyr[o*(nOctaveLayers+) + layer],
Point(c1, r1),
cvRound(SIFT_ORI_RADIUS * scl_octv),
SIFT_ORI_SIG_FCTR * scl_octv,
hist, n);
float mag_thr = (float)(omax * SIFT_ORI_PEAK_RATIO);
for( int j = ; j < n; j++ )
{
int l = j > ? j - : n - ;
int r2 = j < n- ? j + : ; if( hist[j] > hist[l] && hist[j] > hist[r2] && hist[j] >= mag_thr )
{
float bin = j + 0.5f * (hist[l]-hist[r2]) /
(hist[l] - *hist[j] + hist[r2]);
bin = bin < ? n + bin : bin >= n ? bin - n : bin;
kpt.angle = (float)((.f/n) * bin);
keypoints.push_back(kpt);
}
}
}
}
}
}
}

删除边缘效应

除了DoG响应较低的点,还有一些响应较强的点也不是稳定的特征点。DoG对图像中的边缘有较强的响应值,所以落在图像边缘的点也不是稳定的特征点。
一个平坦的DoG响应峰值在横跨边缘的地方有较大的主曲率,而在垂直边缘的地方有较小的主曲率。主曲率可以通过2×2的Hessian矩阵H求出:
【OpenCV】SIFT原理与源码分析:关键点搜索与定位
D值可以通过求临近点差分得到。H的特征值与D的主曲率成正比,具体可参见Harris角点检测算法。
为了避免求具体的值,我们可以通过H将特征值的比例表示出来。令【OpenCV】SIFT原理与源码分析:关键点搜索与定位为最大特征值,【OpenCV】SIFT原理与源码分析:关键点搜索与定位为最小特征值,那么:
【OpenCV】SIFT原理与源码分析:关键点搜索与定位
Tr(H)表示矩阵H的迹,Det(H)表示H的行列式。
【OpenCV】SIFT原理与源码分析:关键点搜索与定位表示最大特征值与最小特征值的比值,则有:
【OpenCV】SIFT原理与源码分析:关键点搜索与定位
上式与两个特征值的比例有关。随着主曲率比值的增加,【OpenCV】SIFT原理与源码分析:关键点搜索与定位也会增加。我们只需要去掉比率大于一定值的特征点。Lowe论文中去掉r=10的点。
// Interpolates a scale-space extremum's location and scale to subpixel
// accuracy to form an image feature. Rejects features with low contrast.
// Based on Section 4 of Lowe's paper.
// 特征点精确定位
static bool adjustLocalExtrema( const vector<Mat>& dog_pyr, KeyPoint& kpt, int octv,
int& layer, int& r, int& c, int nOctaveLayers,
float contrastThreshold, float edgeThreshold, float sigma )
{
const float img_scale = .f/(*SIFT_FIXPT_SCALE);
const float deriv_scale = img_scale*0.5f;
const float second_deriv_scale = img_scale;
const float cross_deriv_scale = img_scale*0.25f; float xi=, xr=, xc=, contr;
int i = ; //三维子像元插值
for( ; i < SIFT_MAX_INTERP_STEPS; i++ )
{
int idx = octv*(nOctaveLayers+) + layer;
const Mat& img = dog_pyr[idx];
const Mat& prev = dog_pyr[idx-];
const Mat& next = dog_pyr[idx+]; Vec3f dD((img.at<short>(r, c+) - img.at<short>(r, c-))*deriv_scale,
(img.at<short>(r+, c) - img.at<short>(r-, c))*deriv_scale,
(next.at<short>(r, c) - prev.at<short>(r, c))*deriv_scale); float v2 = (float)img.at<short>(r, c)*;
float dxx = (img.at<short>(r, c+) +
img.at<short>(r, c-) - v2)*second_deriv_scale;
float dyy = (img.at<short>(r+, c) +
img.at<short>(r-, c) - v2)*second_deriv_scale;
float dss = (next.at<short>(r, c) +
prev.at<short>(r, c) - v2)*second_deriv_scale;
float dxy = (img.at<short>(r+, c+) -
img.at<short>(r+, c-) - img.at<short>(r-, c+) +
img.at<short>(r-, c-))*cross_deriv_scale;
float dxs = (next.at<short>(r, c+) -
next.at<short>(r, c-) - prev.at<short>(r, c+) +
prev.at<short>(r, c-))*cross_deriv_scale;
float dys = (next.at<short>(r+, c) -
next.at<short>(r-, c) - prev.at<short>(r+, c) +
prev.at<short>(r-, c))*cross_deriv_scale; Matx33f H(dxx, dxy, dxs,
dxy, dyy, dys,
dxs, dys, dss); Vec3f X = H.solve(dD, DECOMP_LU); xi = -X[];
xr = -X[];
xc = -X[]; if( std::abs( xi ) < 0.5f && std::abs( xr ) < 0.5f && std::abs( xc ) < 0.5f )
break; //将找到的极值点对应成像素(整数)
c += cvRound( xc );
r += cvRound( xr );
layer += cvRound( xi ); if( layer < || layer > nOctaveLayers ||
c < SIFT_IMG_BORDER || c >= img.cols - SIFT_IMG_BORDER ||
r < SIFT_IMG_BORDER || r >= img.rows - SIFT_IMG_BORDER )
return false;
} /* ensure convergence of interpolation */
// SIFT_MAX_INTERP_STEPS:插值最大步数,避免插值不收敛,程序中默认为5
if( i >= SIFT_MAX_INTERP_STEPS )
return false; {
int idx = octv*(nOctaveLayers+) + layer;
const Mat& img = dog_pyr[idx];
const Mat& prev = dog_pyr[idx-];
const Mat& next = dog_pyr[idx+];
Matx31f dD((img.at<short>(r, c+) - img.at<short>(r, c-))*deriv_scale,
(img.at<short>(r+, c) - img.at<short>(r-, c))*deriv_scale,
(next.at<short>(r, c) - prev.at<short>(r, c))*deriv_scale);
float t = dD.dot(Matx31f(xc, xr, xi)); contr = img.at<short>(r, c)*img_scale + t * 0.5f;
if( std::abs( contr ) * nOctaveLayers < contrastThreshold )
return false; /* principal curvatures are computed using the trace and det of Hessian */
//利用Hessian矩阵的迹和行列式计算主曲率的比值
float v2 = img.at<short>(r, c)*.f;
float dxx = (img.at<short>(r, c+) +
img.at<short>(r, c-) - v2)*second_deriv_scale;
float dyy = (img.at<short>(r+, c) +
img.at<short>(r-, c) - v2)*second_deriv_scale;
float dxy = (img.at<short>(r+, c+) -
img.at<short>(r+, c-) - img.at<short>(r-, c+) +
img.at<short>(r-, c-)) * cross_deriv_scale;
float tr = dxx + dyy;
float det = dxx * dyy - dxy * dxy; //这里edgeThreshold可以在调用SIFT()时输入;
//其实代码中定义了 static const float SIFT_CURV_THR = 10.f 可以直接使用
if( det <= || tr*tr*edgeThreshold >= (edgeThreshold + )*(edgeThreshold + )*det )
return false;
} kpt.pt.x = (c + xc) * ( << octv);
kpt.pt.y = (r + xr) * ( << octv);
kpt.octave = octv + (layer << ) + (cvRound((xi + 0.5)*) << );
kpt.size = sigma*powf(.f, (layer + xi) / nOctaveLayers)*( << octv)*; return true;
}

至此,SIFT第二步就完成了。参见《SIFT原理与源码分析

本文转自:http://blog.csdn.net/xiaowei_cqu/article/details/8087239