【1】np.linalg.eig()函数
输入:计算奇异值和右奇异值的方阵
输出:特征值和特征向量
计算过程:
(1)输入计算的矩阵;(2)矩阵每个维度减去均值;
(3)计算矩阵的协方差矩阵;(4)eig分解
参考代码:
def SVDInArray(sitk_img):
numpy_img=sitk.GetArrayFromImage(sitk_img)
numpy_img[numpy_img<0]=0
sitk_ex=sitk.GetImageFromArray(numpy_img)
rescalFilt = sitk.RescaleIntensityImageFilter()
rescalFilt.SetOutputMaximum(255)
rescalFilt.SetOutputMinimum(0)
itkimage = rescalFilt.Execute(sitk_ex)
numpy_img=sitk.GetArrayFromImage(itkimage)
numpy_img[numpy_img<=0]=0
numpy_img[numpy_img>0]=255
index = np.where(numpy_img==255)
all_point=[]
for i in range(len(index[0])):
one_point_aim=[]
z,y,x=index[0][i],index[1][i],index[2][i]
one_point_aim=[x,y,z]
all_point.append(one_point_aim)
point_aim=np.array(all_point)
m, n = np.shape(point_aim)
col_mean_value=np.mean(point_aim,axis=0)
mean_value=np.tile(col_mean_value, (m, 1))
mean_point_aim1=point_aim-mean_value
covX = np.cov(mean_point_aim1.T)
featValue, featVec= np.linalg.eig(covX)
#特征值未开平方(是svd分解的特征值的平方)
aimpercent=np.double(np.max(featValue)/(featValue[0]+featValue[1]+featValue[2]))
#特征值开平方(与svd分解的特征值相同)
min_aimpercent=np.double(np.sqrt(np.max(featValue))/(np.sqrt(featValue[0])+np.sqrt(featValue[1])+np.sqrt(featValue[2])))
print(featValue)
print(featVec)
print("最大特征值占比:",aimpercent)
print("最大特征值占比(开方):",min_aimpercent)
return aimpercent
【2】np.linalg.svd()函数
矩阵svd分解;可参考numpy.linalg.svd函数_潜艇-CSDN博客
【3】eig和svd分解区别
(1)eig分解只求特征值和特征向量,svd则可以全部求出;
(2)eig求取的特征值是svd取的特征值的平方。
【4】参考文献