[英]Flipping continuous chunks of 1's up to a certain size in a binary numpy matrix
我正在做一個圖像分析項目。 我已經將感興趣的圖片(NxM numpy數組)轉換為二進制格式。 矩陣中的“ 1”是關注區域。 存在感興趣的區域,並且存在無法代表圖像特征的噪點。 例如,在圖像的水平快照中,我不關心孤立的1或2個組,最多5個連續的1。 我想找到一種快速的方法來翻轉它們(即使它們= 0)。
我的MWE用於翻轉孤立的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)
這在我的機器上花費大約9毫秒。 我可以使用長達1秒的例程。
我歡迎任何對代碼的批評(我不是一個優秀的python程序員),以及任何有關如何有效擴展此過程以消除連續的1的孤立小島到通用大小S的小島的想法。
提前致謝
編輯:我意識到國防部是不必要的。 當時,我還想翻轉太小的0島。 一個人可以用== 0代替= mod ....
特定問題的情況
編輯之后,似乎可以使用一些slicing
,從而避免制作中間副本以提高性能。 這是兩行代碼,可實現所需的輸出-
# 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
運行時測試並驗證結果-
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
一般情況
為了使該解決方案通用,可以使用一些信號處理,特別是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)
樣本輸入,輸出-
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]])
聲明:本站的技術帖子網頁,遵循CC BY-SA 4.0協議,如果您需要轉載,請注明本站網址或者原文地址。任何問題請咨詢:yoyou2525@163.com.