如何在NumPy中找到光滑多维数组的局部极小值

Question

假设我有一个NumPy数组，里面存的是一个连续可微函数的评估值，我想找出局部最小值。因为没有噪声，所以每个值比它周围所有邻居的值都小的点，都可以算作局部最小值。

我有一个列表推导式，可以用来处理二维数组，但它忽略了边界上的潜在最小值：

import numpy as N

def local_minima(array2d):
    local_minima = [ index 
                     for index in N.ndindex(array2d.shape)
                     if index[0] > 0
                     if index[1] > 0
                     if index[0] < array2d.shape[0] - 1
                     if index[1] < array2d.shape[1] - 1
                     if array2d[index] < array2d[index[0] - 1, index[1] - 1]
                     if array2d[index] < array2d[index[0] - 1, index[1]]
                     if array2d[index] < array2d[index[0] - 1, index[1] + 1]
                     if array2d[index] < array2d[index[0], index[1] - 1]
                     if array2d[index] < array2d[index[0], index[1] + 1]
                     if array2d[index] < array2d[index[0] + 1, index[1] - 1]
                     if array2d[index] < array2d[index[0] + 1, index[1]]
                     if array2d[index] < array2d[index[0] + 1, index[1] + 1]
                   ]
    return local_minima

不过，这个方法速度比较慢。我还想让它能适用于任意维度的数组。比如，有没有简单的方法可以获取任意维度数组中某个点的所有邻居？或者我是不是在用错方法？我应该使用 numpy.gradient() 吗？

Answer 1

试试这个方法来处理二维数据：

import numpy as N

def local_minima(array2d):
    return ((array2d <= N.roll(array2d,  1, 0)) &
            (array2d <= N.roll(array2d, -1, 0)) &
            (array2d <= N.roll(array2d,  1, 1)) &
            (array2d <= N.roll(array2d, -1, 1)))

这个方法会给你返回一个类似二维数组的结果，里面会标记出局部最小值的位置，标记为真（True）或假（False），这些局部最小值是指它周围四个邻居的值都比它大。

Answer 2

要找到一个任意维度数组中的局部最小值位置，可以使用Ivan的detect_peaks函数，只需要稍微修改一下：

import numpy as np
import scipy.ndimage.filters as filters
import scipy.ndimage.morphology as morphology

def detect_local_minima(arr):
    # https://stackoverflow.com/questions/3684484/peak-detection-in-a-2d-array/3689710#3689710
    """
    Takes an array and detects the troughs using the local maximum filter.
    Returns a boolean mask of the troughs (i.e. 1 when
    the pixel's value is the neighborhood maximum, 0 otherwise)
    """
    # define an connected neighborhood
    # http://www.scipy.org/doc/api_docs/SciPy.ndimage.morphology.html#generate_binary_structure
    neighborhood = morphology.generate_binary_structure(len(arr.shape),2)
    # apply the local minimum filter; all locations of minimum value 
    # in their neighborhood are set to 1
    # http://www.scipy.org/doc/api_docs/SciPy.ndimage.filters.html#minimum_filter
    local_min = (filters.minimum_filter(arr, footprint=neighborhood)==arr)
    # local_min is a mask that contains the peaks we are 
    # looking for, but also the background.
    # In order to isolate the peaks we must remove the background from the mask.
    # 
    # we create the mask of the background
    background = (arr==0)
    # 
    # a little technicality: we must erode the background in order to 
    # successfully subtract it from local_min, otherwise a line will 
    # appear along the background border (artifact of the local minimum filter)
    # http://www.scipy.org/doc/api_docs/SciPy.ndimage.morphology.html#binary_erosion
    eroded_background = morphology.binary_erosion(
        background, structure=neighborhood, border_value=1)
    # 
    # we obtain the final mask, containing only peaks, 
    # by removing the background from the local_min mask
    detected_minima = local_min ^ eroded_background
    return np.where(detected_minima)       

你可以这样使用它：

arr=np.array([[[0,0,0,-1],[0,0,0,0],[0,0,0,0],[0,0,0,0],[-1,0,0,0]],
              [[0,0,0,0],[0,-1,0,0],[0,0,0,0],[0,0,0,-1],[0,0,0,0]]])
local_minima_locations = detect_local_minima(arr)
print(arr)
# [[[ 0  0  0 -1]
#   [ 0  0  0  0]
#   [ 0  0  0  0]
#   [ 0  0  0  0]
#   [-1  0  0  0]]

#  [[ 0  0  0  0]
#   [ 0 -1  0  0]
#   [ 0  0  0  0]
#   [ 0  0  0 -1]
#   [ 0  0  0  0]]]

这段代码的意思是，局部最小值出现在索引 [0,0,3]、[0,4,0]、[1,1,1] 和 [1,3,3] 这些位置：

print(local_minima_locations)
# (array([0, 0, 1, 1]), array([0, 4, 1, 3]), array([3, 0, 1, 3]))
print(arr[local_minima_locations])
# [-1 -1 -1 -1]