When comparing arrays why is "in1d" so much slower than "a==b"

Question

I need to be able to compare two images and extract any unique pixels to create a third image. To accomplish this I did the following:

import cv2
import numpy as np

img = cv2.imread("old.jpg")
img2 = cv2.imread("new.jpg")

image2 = cv2.cvtColor(img2, cv2.COLOR_RGB2RGBA)

for (x,y,z), value in np.ndenumerate(img):
    dif = img[x,y,0] == img2[x,y,0] #only checking one color for speed
    diff = str(dif)
    if "True" in diff:
        image2[x,y,3] = 0
cv2.imwrite("result.png", image2)        

This worked fairly well but the it took around 10 seconds for a 640 x 480 picture and I was hoping to get it closer to about half that time. So I changed this line:

dif = img[x,y,0] == img2[x,y,0]

to

dif = np.in1d(img[x,y,0], img2[x,y,0])

The results are identical but instead of speeding things up it now takes about 3 minutes. I'm at a complete loss as to why.

I realize that iterating over elements in large arrays is going to be time consuming in python but why is in1d so much slower?

(As a side note I would just use the "palette method" but I couldn't see a way to implement it for this purpose due to my limited knowledge of numpy arrays.)

Answer 1

np.in1d checks each element of its first argument against each element of its second element in the worse case. For each element i,j in img it checks if there is an element k,l in img2 with the same value. Meaning that for your 640x480 image you can end up doing (640x480)^2 comparisons.

On the other hand, == checks only elementwise, it checks if element i,j of img is equals to element i,j of img2. It will always has 640x480 comparisons.

np.in1d will work if you have images of different sizes, == will work only for images of the same size.