# Python image triangular mosaic creator

This is a Python script that turns images into "low-poly" art/creates a triangular mosaic from the image, I wrote this script completely by myself.

In short, it takes a filepath to an image, reads the image, and randomly selects a given number of coordinates within the bounding area of the image, and computes the Delaunay triangulation of the coordinates.

It then creates an empty image with the same dimension as the input image, computes the average color of the pixels of the input image within the area defined by each triangle, and puts the triangles with average colors into the new image.

It then applies median filter and gaussian blur to smooth the image, and may convert the image to greyscale.

Sample input (me): Sample output:

4096 points: 8192 points: 16384 points: 8192 points, greyscale: ## Code

import numpy as np
from scipy.spatial import Delaunay
from PIL import Image, ImageFilter

CONDITIONS = [(0, 1, 2), (1, 2, 0), (2, 0, 1)]

def split_triangle(positions):
assert positions.shape == (3, 2)
positions = sorted(positions.tolist(), key=lambda x: x[::-1])
flat = False
for a, b, c in CONDITIONS:
if positions[a] == positions[b]:
base_x1, base_x2 = positions[a], positions[b]
if base_x1 > base_x2:
base_x1, base_x2 = base_x2, base_x1
base_y = positions[a]
tip = positions[c]
flat = True
break

if flat:
return {'triangle': [(base_x1, base_y), (base_x2, base_y), tip], 'flat': True}

else:
low, mid, high = positions
mid_x, mid_y = mid
low_x, low_y = low
high_x, high_y = high
dx, dy = high_x - low_x, high_y - low_y
dy1 = mid_y - low_y
fourth = (low_x + dx * dy1 / dy, mid_y)
if mid_x > fourth:
mid, fourth = fourth, mid

return {'triangles': [[mid, fourth, low], [mid, fourth, high]], 'flat': False}

def scan_triangle(positions):
left_x, right_x = positions, positions
base_y = positions
tip_x, tip_y = positions
h = tip_y - base_y
dx1 = tip_x - left_x
dx2 = tip_x - right_x
step = 1
if tip_y < base_y:
step = -1

points = []

for y in range(round(base_y), round(tip_y+step), step):
ry = round(y)
min_x = round(left_x + dx1 * (y - base_y) / h)
max_x = round(right_x + dx2 * (y - base_y) / h)
points.extend([(x, ry) for x in range(min_x, max_x+1)])

return points

def rasterize_triangle(positions):
data = split_triangle(positions)
if data['flat']:
return scan_triangle(data['triangle'])
else:
return scan_triangle(data['triangles']) + scan_triangle(data['triangles'])

def sample(w, h, n=8192):
points = np.zeros((n, 2))
x = np.random.randint(0, w, n)
y = np.random.randint(0, h, n)
points[:,0] = x
points[:,1] = y
return points

def triangulate_image(path, detail, grey=False):
assert 0 < detail <= 1
image = np.array(Image.open(path))
def triangle_color(x, y):
return np.mean(image[y, x], axis=0)

h, w = image.shape[:2]
num_tiles = w * h / 64
n = round(num_tiles * detail)
samples = sample(w, h, n)
corners = np.array([(0, 0), (0, h - 1), (w - 1, 0), (w - 1, h - 1)])
points = np.concatenate((corners, samples))
triangles = points[Delaunay(points).simplices]
art = np.zeros((h, w, 3), dtype=np.uint8)
for triangle in triangles:
x, y = zip(*rasterize_triangle(triangle))
art[y, x] = triangle_color(x, y)

art = Image.fromarray(art)
art = art.filter(ImageFilter.MedianFilter(3)).filter(ImageFilter.GaussianBlur(1))
if grey:
art = art.convert('L')
return art


The script does exactly what I want, but it is terribly inefficient, how do I improve its performance?

## Edit

Changed the code so that the number of points is relative to level of detail preservation.

Parameter n is replaced by detail, detail is a float number in $$(0, 1]$$, the maximum number of points is reached at detail = 1, the maximum number is width * height / 64, the number of pixel clusters if the image is clustered into squares of 8 pixels, number of points is calculated simply by maximum * detail.

## Update

The theory behind my algorithm is triangle rasterization, for those who doesn't know about it, here is the Wikipedia article about that, and an example implementation, I combined fill_top_triangle and fill_bottom_triangle into one.

• triangulation has a fairly different meaning from yours. You're doing something closer to a triangular mosaic filter. Jun 7 at 12:08
• I measured the time the code spends in each function. Can't really help you improve it, since I think the code is quite hard to read - that may just be me and someone who know the theory behind your algorithm may think differently. Anyways, here's the times: {'sample': 0.0004, 'split_triangle': 0.092, 'scan_triangle': 0.350, 'rasterize_triangle': 0.494, 'triangle_color': 0.646, 'triangulate_image': 1.762}  - times are in seconds. Jun 7 at 14:03
• There actually is a definition of triangulation that matches your use, though I still think your updated title is more clear. Jun 7 at 22:29