在二进制numpy矩阵中将1的连续块翻转到特定大小

我正在进行一个图像分析项目。 我把我感兴趣的图片(一个NxM numpy数组)变成了二进制格式。 矩阵中的“1”是感兴趣的区域。 存在感兴趣的区域,并且存在不可能表示图像上的特征的噪声。 例如,在图像的水平快照中,隔离的1个或2个组,例如5个连续的1个对我来说不感兴趣。 我想找到一种快速翻转它们的方法(即使它们= 0)。

我的MWE用于翻转孤立的1:

import numpy as np
img = np.random.choice([0,1],size=(1000,1000), p=[1./2,1./2])

#now we take the second derivative of the matrix in the horizontal axis
#since we have a binary matrix, an isolated 1, that is [...010...] is captured
#by a second derivative entry equal to -2
#because ([...010...]->dx->[...1,-1,...]->dx->[...-2...]

ddx_img = np.diff(np.diff(img,1),1)
to_flip = np.where(ddx_img==-2) #returns a tuple of [x,y] matrix entries

# the second derivative eats up an index position on horizontally, so I need to add
# +1 to the horizontal axis of the tuple

temp_copy = to_flip[1].copy() #cannot modify tuple directly, for some reason its read only
temp_copy+=1
to_flip = (to_flip[0],temp_copy)

#now we can flip the entries by adding +1 to the entries to flip and taking mod 2
img[to_flip]=mod(img[to_flip]+1,2)

这在我的机器上大约需要9毫秒。 我可以做最多1秒的例行程序。

我欢迎任何对代码的批评(我不是一个优秀的python程序员),以及关于如何有效地扩展此过程以消除连续1的孤岛以及通用大小S岛的任何想法。

提前致谢

编辑:我意识到mod是不必要的。 在我这样做的时候,我还想翻转一个0的小岛。 可以将= mod ....替换为= 0

采纳答案:

特定问题的案例

在编辑之后,似乎你可以使用一些slicing ,从而避免制作中间副本以提高性能。 这是用于实现所需输出的两行代码 -

# Calculate second derivative
ddx_img = np.diff(np.diff(img,1),1)

# Get sliced version of img excluding the first and last columns 
# and use mask with ddx elements as "-2" to zeros
img[:,1:-1][ddx_img==-2] = 0

运行时测试并验证结果 -

In [42]: A = np.random.choice([0,1],size=(1000,1000), p=[1./2,1./2])

In [43]: def slicing_based(A):
    ...:    img = A.copy()
    ...:    ddx_img = np.diff(np.diff(img,1),1)
    ...:    img[:,1:-1][ddx_img==-2] = 0
    ...:    return img
    ...: 
    ...: 
    ...: def original_approach(A):
    ...: 
    ...:    img = A.copy()
    ...: 
    ...:    ddx_img = np.diff(np.diff(img,1),1)
    ...:    to_flip = np.where(ddx_img==-2)
    ...: 
    ...:    temp_copy = to_flip[1].copy()
    ...:    temp_copy+=1
    ...:    to_flip = (to_flip[0],temp_copy)
    ...: 
    ...:    img[to_flip] = 0
    ...: 
    ...:    return img
    ...: 

In [44]: %timeit slicing_based(A)
100 loops, best of 3: 15.3 ms per loop

In [45]: %timeit original_approach(A)
10 loops, best of 3: 20.1 ms per loop

In [46]: np.allclose(slicing_based(A),original_approach(A))
Out[46]: True

通用案例

为了使解决方案通用,可以使用一些信号处理,特别是2D convolution ,如下所示 -

# Define kernel
K1 = np.array([[0,1,1,0]]) # Edit this for different island lengths
K2 = 1-K1

# Generate masks of same shape as img amd based on TRUE and inverted versions of 
# kernels being convolved and those convolved sums being compared against the 
# kernel sums indicating those spefic positions have fulfiled both the ONES 
# and ZEROS criteria
mask1 = convolve2d(img, K1, boundary='fill',fillvalue=0, mode='same')==K1.sum()
mask2 = convolve2d(img==0, K2, boundary='fill',fillvalue=0, mode='same')==K2.sum()

# Use a combined mask to create that expanses through the kernel length 
# and use it to set those in img to zeros
K3 = np.ones((1,K1.size))
mask3 = convolve2d(mask1 & mask2, K3, boundary='fill',fillvalue=0, mode='same')>0
img_out = img*(~mask3)

样本输入,输出 -

In [250]: img
Out[250]: 
array([[0, 1, 1, 1, 0, 1, 1, 1],
       [1, 1, 1, 1, 1, 1, 0, 1],
       [1, 0, 1, 1, 1, 1, 0, 0],
       [1, 1, 1, 1, 0, 1, 0, 1],
       [1, 1, 0, 1, 1, 0, 1, 1],
       [1, 0, 1, 1, 1, 1, 1, 1],
       [1, 1, 0, 1, 1, 0, 1, 0],
       [1, 1, 1, 0, 1, 1, 1, 1]])

In [251]: img_out
Out[251]: 
array([[0, 1, 1, 1, 0, 1, 1, 1],
       [1, 1, 1, 1, 1, 1, 0, 1],
       [1, 0, 1, 1, 1, 1, 0, 0],
       [1, 1, 1, 1, 0, 1, 0, 1],
       [1, 1, 0, 0, 0, 0, 0, 1],
       [1, 0, 1, 1, 1, 1, 1, 1],
       [1, 1, 0, 0, 0, 0, 0, 0],
       [1, 1, 1, 0, 1, 1, 1, 1]])
author: divakar

参考更多解答: Flipping continuous chunks of 1's up to a certain size in a binary numpy matrix ,转载请保留出处在二进制numpy矩阵中将1的连续块翻转到特定大小及作者信息

Statement: We respect knowledge and authors. Since the content comes from the Internet and is intended for scientific research, any reprinters should retain the author's signature and origin. If you are the author of the content and feel in dispute, please contact email: 1076545519@qq.com. We will find out the situation and deal with it in time. We sincerely thank the author for his hard work.


更多:binary-data