1

I need to segment out anomalies in a greyscale image. In a certain place in my algorithm, I compute a matrix that contains the known pixel intensities that I need to set to zero. How would I do this?

For example:
The computed pixel intensities: (array([ 94, 95, 96, 97, 98, 99, 100, 101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 111, 112, 113, 114, 115, 116, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143, 144, 145, 146, 147, 148, 149, 150, 151]),)

The picture is of size (480,640) :
Printed it gives:
[[ 86 90 97 ..., 142 152 157]
[ 85 89 97 ..., 145 154 158]
[ 83 87 95 ..., 154 158 159]
...,
[130 134 139 ..., 156 154 154]
[130 134 140 ..., 154 153 152]
[130 134 141 ..., 154 153 152]]

I realize that for each pixel I could go through the intensity matrix. This would, however, be too expensive. NumPy experts I need your help!

Improve this question
1
  • The pixel intensities that I want to set to zero are computed on the flow and can be different. Commented Feb 16, 2015 at 12:53

2 Answers 2

Reset to default
3

Simply

oI[ (oI >93) & (oI < 152)  ] = 0

should do I think...

According to your question, you want to have specific non-contiguous numbers in an array. That is best handled by a map-reduce algorithm. Let us say you want the following numbers to be set to 0: 

numList = np.array([2, 15, 100, 56])

Then you can set a mask like so, and reduce them to a single mask:

mask     = (oI == 2)|(oI == 15)|(oI == 100)|(oI == 56)
oI[mask] = 0

This is clearly not a good solution for a big list of numbers. So you can do ...

mask = reduce( lambda m, n: m|n ,[ oI == i for i in numList]) 
oI[mask] = 0

If it is contiguous, like between a and b, then you just use the earlier approach ...

mask = (oI > a) & (oI < b)

Of course you can combine the masks in any set of contiguous and non contiguous sets of numbers. Just | them.

Hope this helps ...

Improve this answer
Sign up to request clarification or add additional context in comments.

2 Comments

The pixel intensities that I want to set to zero are computed on the flow and can be different. Can I maybe come up with conditions based on a vector of numbers. Say more than 91 and less than 151?
Thank you for your edit. My values will be contiguous. But I might receive two contiguous lists. Say [3,4,5,6,7,90,91,92,93,94]. How would I find out the endpoints of these contiguous values ([a = 1, b =7][a = 90, b = 94] in this example).
1

To set to zero all pixels in an image array which have values from 91 to 151, inclusive, use:

import numpy as np
newimage = np.where(np.logical_and(91<=oldimage, oldimage<=151), 0, oldimage)

To set to zero all pixels in an image array whose values belong to some array vct, use:

newimage = np.where(np.in1d(oldimage, vct).reshape(oldimage.shape), 0, oldimage)

Example

Suppose we have an oldimage like this:

In [12]: oldimage
Out[12]: 
array([[0, 1, 2],
       [3, 4, 5]])

And we have a list of numbers called vct:

In [13]: vct
Out[13]: array([3, 5])

Let's set to zero all pixels in oldimage that are also in vct:

In [14]: newimage = np.where(np.in1d(oldimage, vct).reshape(oldimage.shape), 0, oldimage)

In [15]: newimage
Out[15]: 
array([[0, 1, 2],
       [0, 4, 0]])
Improve this answer

4 Comments

The pixel intensities that I want to set to zero are computed on the flow and can be different. Can I maybe come up with conditions based on a vector of numbers. Say more than 91 and less than 151?
@Talha I just added to the answer a version that handles a condition based on an arbitrary vector of numbers.
Thank you so much for your response! I had actually proceeded with writing a function with a for loop that would find bounds but this seems like it is the vectored solution. I will check which algorithm is faster and by how much in an hour or so.
Turns out the in1d function doesn't scale so well with large arrays. With the in1d function I get about 6 FPS. With my bound finding algorithm and the threshold method I get about 90 FPS.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.