Numpy最快的3D到2D投影

时间:2015-03-12 11:24:52

标签: python arrays optimization numpy

我有二维数据的3D数组。我想把它投射到3个2D图像 - 侧面,正面,鸟眼。

我写了代码:

for x in range(data.shape[2]):
    for y in range(data.shape[0]):
        val = 0
        for z in range(data.shape[1]):
            if data[y][z][x] > 0:
                val = 255
                break
        side[y][x] = val

但对于~700x300x300矩阵来说,这是非常缓慢的(75秒!)。

实现此任务的最快方法是什么?

编辑:

为了保存图像,我使用了:

sideImage = Image.fromarray(side)
sideImage.convert('RGB').save("sideImage.png")

3 个答案:

答案 0 :(得分:5)

当我拥有3D数据时,我倾向于将其视为一个多维数据集'包含2D图像的行,列和切片(或面板)。每个切片或面板都是尺寸为(rows, cols)的2D图像。我通常会这样想:

3D data cube

(0,0,0)位于切片的左上角。使用numpy索引,只需选择您对感兴趣的3D数组部分而不编写自己的循环非常容易:

>>> import numpy as np
>>> import matplotlib.pyplot as plt
>>> np.set_printoptions(precision=2)

# Generate a 3D 'cube' of data
>>> data3D = np.random.uniform(0,10, 2*3*5).reshape((2,3,5))
>>> data3D
array([[[ 7.44,  1.14,  2.5 ,  3.3 ,  6.05],
        [ 1.53,  8.91,  1.63,  8.95,  2.46],
        [ 3.57,  3.29,  6.43,  8.81,  6.43]],

       [[ 4.67,  2.67,  5.29,  7.69,  7.59],
        [ 0.26,  2.88,  7.58,  3.27,  4.55],
        [ 5.84,  9.04,  7.16,  9.18,  5.68]]])

# Grab some "views" of the data
>>> front  = data3D[:,:,0]  # all rows and columns, first slice
>>> back   = data3D[:,:,-1] # all rows and cols, last slice
>>> top    = data3D[0,:,:]  # first row, all cols, all slices 
>>> bottom = data3D[-1,:,:] # last row, all cols, all slices
>>> r_side = data3D[:,-1,:] # all rows, last column, all slices
>>> l_side = data3D[:,0,:]  # all rows, first column, all slices

看看前面的样子:

>>> plt.imshow(front, interpolation='none')
>>> plt.show()

front of data cube

答案 1 :(得分:1)

您可以按如下方式计算:

>>> data = np.random.random_sample((200, 300, 100)) > 0.5
>>> data.any(axis=-1).shape # show the result has the shape we want
(200, 300)
>>> data.any(axis=-1)
array([[ True,  True,  True, ...,  True,  True,  True],
       [ True,  True,  True, ...,  True,  True,  True],
       [ True,  True,  True, ...,  True,  True,  True],
       ...,
       [ True,  True,  True, ...,  True,  True,  True],
       [ True,  True,  True, ...,  True,  True,  True],
       [ True,  True,  True, ...,  True,  True,  True]], dtype=bool)
>>>

如果需要,您可以缩放值

>>> data.any(axis=-1) * 255
array([[255, 255, 255, ..., 255, 255, 255],
       [255, 255, 255, ..., 255, 255, 255],
       [255, 255, 255, ..., 255, 255, 255],
       ...,
       [255, 255, 255, ..., 255, 255, 255],
       [255, 255, 255, ..., 255, 255, 255],
       [255, 255, 255, ..., 255, 255, 255]])
>>>

答案 2 :(得分:1)

一段时间后,我将下面的内容写成3D阵列的可视化辅助工具。也是一个很好的学习练习。

# Python 2.7.10
from __future__ import print_function
from numpy import *

def f_Print3dArray(a_Array):
    v_Spacing = (len(str(amax(abs(a_Array)))) + 1) if amin(a_Array)\
        < 0 else (len(str(amax(a_Array))) + 1)
    for i in a_Array[:,:,::-1].transpose(0,2,1):
        for index, j in enumerate(i):
            print(" " * (len(i) - 1 - index) + "/ ", end="")
            for k in j:
                print(str(k).ljust( v_Spacing + 1), end="")
            print("/")
        print()

a_Array = arange(27).reshape(3, 3, 3)
print(a_Array)
print()

f_Print3dArray(a_Array)

转换这个:

[[[ 0  1  2]
  [ 3  4  5]
  [ 6  7  8]]

 [[ 9 10 11]
  [12 13 14]
  [15 16 17]]

 [[18 19 20]
  [21 22 23]
  [24 25 26]]]

对此:

  / 2   5   8   /
 / 1   4   7   /
/ 0   3   6   /

  / 11  14  17  /
 / 10  13  16  /
/ 9   12  15  /

  / 20  23  26  /
 / 19  22  25  /
/ 18  21  24  /

希望它有所帮助。