I find your numpy code to be quite rational as far as my skills go, but your script have several things to improve on besides that.

Improvable in current scope :

• Removing duplicate code thanks to a helper function
• Painting the whole background to simplify
• Removing (yet) unecessary line w, h, k = data.shape
• Naming error, or logic error ? Background and edge color seem swapped. Changing <= level to > level
• Some non-compliant and unexplicit variable names

Code showing this

# -*- coding: utf-8 -*-
import numpy as np
from PIL import Image

def get_dist_array(data, shifted_data, level):
    distance_axis = np.absolute(np.sum(data - shifted_data, axis=2))
    return np.asarray(distance_axis > level, dtype="int32")

def stroke(image_path, output, level=80, edge_color=[255,255,255], blackground_color=[0,0,0]):
    img = Image.open(image_path)
    data = np.asarray(img, dtype="int32")
    level = min(max(1,level), 255)
    edges_right = get_dist_array(data, np.concatenate((data[:, 1:], data[:,-1:]), axis=1), level)
    edges_down = get_dist_array(data, np.concatenate((data[1:,:], data[-1:,:]), axis=0), level)
    data[:] = blackground_color
    edges = (edges_right+edges_down) > 1
    data[edges] = edge_color
    img = Image.fromarray(np.asarray(np.clip(data, 0, 255), dtype="uint8"), "RGB")
    img.save(output)

if __name__ == "__main__":
    stroke("images/bob.png", "new_test.jpg")

Other things to consider worth improving :

• Get rid of hardcoded file names. You could use argparse here
• RGBA support