Using Python to create a new Numpy array from an existing one

Question

I need to use Python and Numpy to take a grayscale image (as a Numpy array), then iterate through it by pixel in order to differentiate the image in the X a direction. I can't use any functions to do this automatically, I need to iterate unfortunately. I need to use the derivative: F(x, y) = F(x, y+1) - F(x, y) to each pixel and return the output in the form of a new image (in a Numpy array).

A simple 4 pixel example 10 15 5 25 Would return 5 10 20

I want to take the absolute value (eliminating negative values) and have the output width be 1 column shorter than the input (since no calculation can be performed on the last column).

I can read the array using np.nditer, but I could really use some help figuring out how to apply that calculation to each element and return the results.

I figured this out in Java fairly quickly, can any of the Python gurus out there help me convert this?

public class ArrayTest {
public static void main(String[] args) {

    int[] arrayOne = { 5, 10, 20, 5 };
    int[] newArray = new int[arrayOne.length];

    for (int i = 0; i < arrayOne.length - 1; i++) {

        newArray[i] = Math.abs(arrayOne[i + 1] - arrayOne[i]);

        System.out.println(newArray[i]);
    }
}

}

Its second-order accurate rather than your requested first-order, but you could use np.gradient — tmdavison
– tmdavison, Commented Sep 16, 2015 at 14:06
I wish I could use these functions, but the requirements specify that I 'feel the pain' and iterate through each pixel. For example, Pixel a is 5 and Pixel b is 10, I need to create a new Pixel a of 5 and add to a new array...and so on. — yodish
– yodish, Commented Sep 16, 2015 at 14:13

areuexperienced · Accepted Answer · 2015-09-16 21:06:19Z

3

Following my snide comment about homework:

Looking at your Java code, I think this is what you want?

import numpy as np

data = np.array([10, 15, 5, 25])
diff = np.abs(data[:-1] - data[1:])

print diff
array([ 5, 10, 20])

EDIT:

I'm simply taking every value of the array apart from the last one (as there's nothing to be computed there) and differencing that versus every value apart from the first.

print data[:-1]
array([10, 15,  5])
print data[1:]
array([15,  5, 25])

data[1:] - data[:-1] is equivalent to F(x) = F(x+1) - F(x).

I imagine you're familiar with slicing notation from using lists.

Using a loop:

new = np.empty(shape = data.shape[0]-1)

for i in range(0, new.shape[0]):
    new[i] = np.abs(data[i+1] - data[i])

As @Joe Kington says, you generally shouldn't need to do this, as numpy allows using vectorised expressions (operations computed on the entire array rather than for each element), which makes the code a lot faster. Not a requirement in this trivial example, but would probably give you noticeable benefits if you're doing this with lots of large arrays.

EDIT 2:

Using a loop in the 2D case:

import numpy as np
data =  np.array([10, 15, 5, 25])
data_2d = np.repeat(data,2).reshape(-1,2) #make some 2d data
data_2d[:,1] = data_2d[:,1] + 100 #make the y axis different so we can tell them apart easier

print data_2d
[[ 10 110]
 [ 15 115]
 [  5 105]
 [ 25 125]]

'''
Making a new array to store the results, copying over the Y values.
The X values we will change later. Note that not using the .copy() 
method would create a VIEW of data_2d, so when we change new,
data_2d would change as well.
'''

new = data_2d[:-1,:].copy()

print new.shape
(3,2) # 3 here is the number of elements per axis, 2 is the number of axes. 

for i in range(0,data_2d.shape[0]-1): # looping the X axis
   new[i,0] = np.abs(data_2d[i+1,0] - data_2d[i,0]) # referencing the X axis explicitly

print new
[[  5 110]
 [ 10 115]
 [ 20 105]]

edited Sep 16, 2015 at 21:06

answered Sep 16, 2015 at 14:56

areuexperienced

2,0713 gold badges18 silver badges28 bronze badges

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

13 Comments

yodish Over a year ago

I appreciate the help. If that is returning 5,10,20 then that is basically what i'm looking for, although I don't quite understand it yet. Does a for loop not need to be included? In my limited experience with python, I would iterate the array with np.nditer and follow that with a for loop.

Joe Kington Over a year ago

@yodish - You can use a loop, but its typically an anti-pattern when dealing with numpy arrays. It will be much faster to use vectorized expressions. Regardless, you wouldn't want np.nditer in this case. If you were going to iterate through, use indexing just like you would in your Java example.

areuexperienced Over a year ago

@yodish - I edited my answer with some explanation, plus an example of how this'd look like using a loop. I hope this helps.

yodish Over a year ago

Thank you for breaking that down, it's making a little more sense to me. So, in my case, where i'm taking an 'image' (a numpy array), performing calculations, and returning 'new_image' (another numpy array); using your loop example, would I just replace 'data' with 'image' and 'new' with 'new_image'? I apologize for not testing this myself, I don't have python installed on my work machine and admin rights are so elusive...

areuexperienced Over a year ago

It'd be in 2D so new_image would need to be a 2D array and you'd need to refer to the values on the X axis explicitly when you're looping over it.

|

Collectives™ on Stack Overflow

Using Python to create a new Numpy array from an existing one

1 Answer 1

13 Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

13 Comments

Your Answer

Sign up or log in

Post as a guest

Related