python m和n的运算_快速计算图像（M，N）和模板（3，3）之间差平方和的快速方法？...

你可以用as_strided函数与numpy的广播相结合来做一些惊人的事情。以下是您的函数的两个版本：import numpy as np

from numpy.lib.stride_tricks import as_strided

def sumsqdiff(input_image, template, valid_mask=None):

if valid_mask is None:

valid_mask = np.ones_like(template)

total_weight = valid_mask.sum()

window_size = template.shape

ssd = np.empty((input_image.shape[0] - window_size[0] + 1,

input_image.shape[1] - window_size[1] + 1))

for i in xrange(ssd.shape[0]):

for j in xrange(ssd.shape[1]):

sample = input_image[i:i + window_size[0], j:j + window_size[1]]

dist = (template - sample) ** 2

ssd[i, j] = (dist * valid_mask).sum()

return ssd

def sumsqdiff2(input_image, template, valid_mask=None):

if valid_mask is None:

valid_mask = np.ones_like(template)

total_weight = valid_mask.sum()

window_size = template.shape

# Create a 4-D array y, such that y[i,j,:,:] is the 2-D window

# input_image[i:i+window_size[0], j:j+window_size[1]]

y = as_strided(input_image,

shape=(input_image.shape[0] - window_size[0] + 1,

input_image.shape[1] - window_size[1] + 1,) +

window_size,

strides=input_image.strides * 2)

# Compute the sum of squared differences using broadcasting.

ssd = ((y - template) ** 2 * valid_mask).sum(axis=-1).sum(axis=-1)

return ssd

这里有一个ipython会话来比较它们。在

我将用于演示的模板：

^{pr2}$

我们可以检查结果的小输入：In [73]: x

Out[73]:

array([[ 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., 27.],

[ 28., 29., 30., 31., 32., 33., 34.]])

将这两个函数应用于x，并检查是否得到相同的结果：In [74]: sumsqdiff(x, template)

Out[74]:

array([[ 856., 1005., 1172., 1357., 1560.],

[ 2277., 2552., 2845., 3156., 3485.],

[ 4580., 4981., 5400., 5837., 6292.]])

In [75]: sumsqdiff2(x, template)

Out[75]:

array([[ 856., 1005., 1172., 1357., 1560.],

[ 2277., 2552., 2845., 3156., 3485.],

[ 4580., 4981., 5400., 5837., 6292.]])

现在输入更大的“图像”：In [76]: z = np.random.randn(500, 500)

并检查性能：In [77]: %timeit sumsqdiff(z, template)

1 loops, best of 3: 3.55 s per loop

In [78]: %timeit sumsqdiff2(z, template)

10 loops, best of 3: 33 ms per loop

不太寒酸。：)

两个缺点：sumsqdiff2中的计算将生成一个临时数组，对于3x3模板，该数组的大小是input_image的9倍。(通常它是template.size倍于input_image的大小。)

当你对代码进行计算机化处理时，这些“跨步技巧”不会对你有所帮助。当转换为Cython时，您通常会回到使用numpy进行矢量化时消除的循环。在