Here is my code for edge detection using the Sobel operator:
from PIL import Image
import numpy as np
from scipy import misc
a = np.array([1, 2, 1])
b = np.array([1, 0, -1])
Gy = np.outer(a, b)
Gx = np.rot90(Gy, k=3)
def apply(X):
a = (X * Gx)
b = (X * Gy)
return np.abs(a.sum()) + np.abs(b.sum())
data = np.uint8(misc.lena())
data2 = np.copy(data)
center = offset = 1
for i in range(offset, data.shape[0]-offset):
for j in range(offset, data.shape[1]-offset):
X = data[i-offset:i+offset+1, j-offset:j+offset+1]
data[i, j] = apply(X)
image = Image.fromarray(data)
image.show()
image = Image.fromarray(data2)
image.show()
Which results in:
Instead of:
For what it's worth, I'm fairly certain my for loops and general idea of image kernels are correct. For example, I was able to generate this custom filter (Gaussian with center subtracted):
What's wrong with my Sobel filter?
Finally figured it out. I shouldn't have been modifying the array in place because it obviously changes the values that are computed in subsequent applications of the filter. This works:
Which produces: