当前位置：首页 > news >正文

opencv矩阵转eigen_numpy opencv matlab eigen SVD结果对比

news 来源：原创 2024/5/17 12:26:33

参考

https://zhuanlan.zhihu.com/p/26306568

https://byjiang.com/2017/11/18/SVD/

http://www.bluebit.gr/matrix-calculator/

https://stackoverflow.com/questions/3856072/single-value-decomposition-implementation-c

https://stackoverflow.com/questions/35665090/svd-matlab-implementation

矩阵奇异值分解简介及C++/OpenCV/Eigen的三种实现

https://blog.csdn.net/fengbingchun/article/details/72853757

numpy.linalg.svd 源码

https://docs.scipy.org/doc/numpy/reference/generated/numpy.linalg.svd.html

计算矩阵的特征值与特征向量的方法

https://blog.csdn.net/Junerror/article/details/80222540

https://jingyan.baidu.com/article/27fa7326afb4c146f8271ff3.html

不同的库计算结果不一致

原因在于特征向量不唯一，特征值是唯一的

来源

https://stackoverflow.com/questions/35665090/svd-matlab-implementation

Both are correct... The rows of the v you got from numpy are the eigenvectors of M.dot(M.T) (the transpose would be a conjugate transpose in the complex case). Eigenvectors are in the general case defined only up to a multiplicative constant, so you could multiply any row of v by a different number, and it will still be an eigenvector matrix.

There is the additional constraint on v that it be a unitary matrix, which loosely translates to its rows being orthonormal. This reduces your available choices for every eigenvector to only 2: the normalized eigenvector pointing in either direction. But you still get to multiply any row by -1 and still have a valid v.

A = U * S * V

1 手动计算

给定一个大小为

的矩阵

，虽然这个矩阵是方阵，但却不是对称矩阵，我们来看看它的奇异值分解是怎样的。

由

进行对称对角化分解，得到特征值为

，

，相应地，特征向量为

，

；由

进行对称对角化分解，得到特征值为

，

当 lamda1 = 32

AA.T - lamda1*E = [-7 7

7 -7]

线性变换【-1 1

0 0】

-x1 + x2 = 0

x1 = 1 x2 = 1 特征向量为【1 1】.T 归一化为【1/sqrt(2), 1/sqrt(2)】

x1 = -1 x2 = -1 特征向量为【-1 -1】.T 归一化为【-1/sqrt(2), -1/sqrt(2)】

当 lamda2 = 18

AA.T - lamda2*E = [7 7

7 7]

线性变换【1 1

0 0】

x1 + x2 = 0

如果x1 = -1 x2 = 1 特征向量为【-1 1】.T 归一化为【-1/sqrt(2), 1/sqrt(2)】

如果 x1 = 1 x2 = -1 特征向量为【-1 1】.T 归一化为【1/sqrt(2), -1/sqrt(2)】可见特征向量不唯一

特征向量为

，

。取

，则矩阵

的奇异值分解为

2 MATLAB 结果与手动计算不同

AB = [[ 4 4],

[-3 3]]

[U,S,V]=svd(AB);

V'#需要转置

AB =

4 4

-3 3

U =

-1 0

0 1

S =

5.6569 0

0 4.2426

V =

-0.7071 -0.7071

-0.7071 0.7071

3 NUMPY结果与手动计算不同, 与matlab相同它们都是调用lapack的svd分解算法。

importnumpy as npimportmathimportcv2

a= np.array([[4,4],[-3,3]])#a = np.random.rand(2,2) * 10

print(a)

u, d, v=np.linalg.svd(a)print(u)print(d)print(v)#不需要转置

A = [[ 4 4]

[-3 3]]

U =

[[-1. 0.]

[ 0. 1.]]

[5.65685425 4.24264069]

[[-0.70710678 -0.70710678]

[-0.70710678 0.70710678]]

4 opencv结果与手动计算结果相同

importnumpy as npimportmathimportcv2

a= np.array([[4,4],[-3,3]],dtype=np.float32)#a = np.random.rand(2,2) * 10

print(a)

S1, U1, V1=cv2.SVDecomp(a)print(U1)print(S1)print(V1)#不需要转置

a = [[ 4. 4.]

[-3. 3.]]

U =

[[0.99999994 0. ]

[0. 1. ]]

S = [[5.656854 ]

[4.2426405]]

V =

[[ 0.70710677 0.70710677]

[-0.70710677 0.70710677]]

5 eigen结果与手动计算相同

#include #include #include #include #include #include "opencv2/imgproc/imgproc.hpp"#include

using namespace std;

using namespace cv;//using Eigen::MatrixXf;

using namespace Eigen;

using namespace Eigen::internal;

using namespace Eigen::Architecture;

int GetEigenSVD(Mat&Amat, Mat &Umat, Mat &Smat, Mat &Vmat)

{//-------------------------------svd测试 eigen

Matrix2f A;

A(0,0)=Amat.at(0,0);

A(0,1)=Amat.at(0,1);

A(1,0)=Amat.at(1,0);

A(1,1)=Amat.at(1,1);

JacobiSVD<:matrixxf> svd(A, ComputeThinU |ComputeThinV );

Matrix2f V= svd.matrixV(), U =svd.matrixU();

Matrix2f S= U.inverse() * A * V.transpose().inverse(); // S = U^-1 * A * VT * -1

//Matrix2f Arestore = U * S *V.transpose();// printeEigenMat(A ,"/sdcard/220/Ae.txt");// printeEigenMat(U ,"/sdcard/220/Ue.txt");// printeEigenMat(S ,"sdcard/220/Se.txt");// printeEigenMat(V ,"sdcard/220/Ve.txt");// printeEigenMat(U * S * V.transpose() ,"sdcard/220/U*S*VTe.txt");

Umat.at(0,0) =U(0,0);

Umat.at(0,1) = U(0,1);

Umat.at(1,0) = U(1,0);

Umat.at(1,1) = U(1,1);

Vmat.at(0,0) =V(0,0);

Vmat.at(0,1) = V(0,1);

Vmat.at(1,0) = V(1,0);

Vmat.at(1,1) = V(1,1);

Smat.at(0,0) =S(0,0);

Smat.at(0,1) = S(0,1);

Smat.at(1,0) = S(1,0);

Smat.at(1,1) = S(1,1);// Smat.at(0,0) =S(0,0);// Smat.at(0,1) = S(1,1);//-------------------------------svd测试 eigenreturn0;