缩小时的值插值 [英] Interpolation of values when zooming down

问题描述

我有一个2D数组，我想下采样以将其与另一个数组进行比较.

I have a 2D array that I would like to down sample to compare it to another.

假设我的数组x是512x512，我想要一个数组y 128x128，其中y的元素是使用x的overs 4x4块的插值构建的>(此插值可能只是取平均值，但其他方法(例如几何平均值)可能会很有趣)

Lets say my array x is 512x512, I'd like an array y 128x128 where the elements of y are build using an interpolation of the values overs 4x4 blocks of x (this interpolation could just be taking the average, but other methodes, like geometric average, could be interesting)

到目前为止，我查看了scipy.ndimage.interpolation.zoom，但没有得到想要的结果

So far I looked at scipy.ndimage.interpolation.zoom but I don't get the results I want

>> x = np.arange(16).reshape(4,4)
>> print(x)
[[ 0  1  2  3]
 [ 4  5  6  7]
 [ 8  9 10 11]
 [12 13 14 15]]
>> y = scipy.ndimage.interpolation.zoom(x, 0.5)
>> print(y)
[[ 0  3]
 [12 15]]

我希望y是

[[ 2.5  4.5]
 [10.5 12.5]]

请注意，仅设置dtype=np.float32并不能解决问题...

Note that simply setting dtype=np.float32 doesn't solve that ...

推荐答案

您似乎要查找的是4块的均值，而zoom无法获得，因为zoom使用内插法(请参见docstring)

What you seem to be looking for is the mean over blocks of 4, which is not obtainable with zoom, since zoom uses interpolation (see its docstring)

要获取显示的内容，请尝试以下

To obtain what you show, try the following

import numpy as np
x = np.arange(16).reshape(4, 4)

xx = x.reshape(len(x) // 2, 2, x.shape[1] // 2, 2).transpose(0, 2, 1, 3).reshape(len(x) // 2, x.shape[1] // 2, -1).mean(-1)

print xx

这产生

[[  2.5   4.5]
 [ 10.5  12.5]]

或者，也可以使用sklearn.feature_extraction.image.extract_patches

from sklearn.feature_extraction.image import extract_patches

patches = extract_patches(x, patch_shape=(2, 2), extraction_step=(2, 2))

xx = patches.mean(-1).mean(-1)

print xx

但是，如果您的目标是以优美的方式对图像进行二次采样，则对图像的块进行均值并不是正确的方法:很可能会导致混叠效果.在这种情况下，您应该使用scipy.ndimage.gaussian_filter(例如sigma=0.35 * subsample_factor)稍微平滑图像，然后只需通过索引[::2, ::2]

However, if your goal is to subsample an image in a graceful way, then taking the mean over blocks of the image is not the right way to do it: It is likely to cause aliasing effects. What you should do in this case is smooth the image ever so slightly using scipy.ndimage.gaussian_filter (e.g. sigma=0.35 * subsample_factor) and then subsample simply by indexing [::2, ::2]

这篇关于缩小时的值插值的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！