I am trying to make an image mosaic generator using pyvips. So basically, given an image (called original in the following) create a new, bigger, image that resembles the original one except each pixel (or more realistically groups of pixels) are replaced by smaller distinct image tiles.
I was drawn to pyvips because it is said it can handle huge images and that it can process images without having to load them completely into memory.
However, I am having an issue creating a blank mosaic to then populate with tile images.
In the code below I try joining tiles together row by row to create a mosaic but this code unfortunately eats through my RAM and always segfaults.
import os
import pyvips
from os.path import join
from scipy.spatial import cKDTree
class Mosaic(object):
def __init__(self, dir_path, original_path, tree=None, averages=None):
self.dir_path = dir_path
self.original = original_path
self.tree = tree
if averages:
self.averages = averages
else:
self.averages = {}
def get_image(self, path):
return pyvips.Image.new_from_file(path, access="sequential")
def build_tree(self):
for root, dirs, files in os.walk(self.dir_path):
print('Loading images from', root, '...')
for file_name in files:
path = join(root, file_name)
try:
image = pyvips.Image.new_from_file(path)
self.averages[self.avg_rgb(image)] = path
except pyvips.error.Error:
print('File', path, 'not recognized as an image.')
self.tree = cKDTree(self.averages.keys())
print('Loaded', len(self.averages), 'images.')
def avg_rgb(self, image):
m = image.stats()
return tuple(m(4,i)[0] for i in range(1,4))
def get_tile_name(self, patch):
avg = self.avg_rgb(patch)
index = self.tree.query(avg)[1]
return self.averages[tuple(self.tree.data[index])]
def get_tile(self, x, y, step):
patch = self.get_image(self.original).crop(x, y, step, step)
patch_name = self.get_tile_name(patch)
return pyvips.Image.new_from_file(patch_name, access="sequential")
def make_mosaic(self, tile_num, tile_size, mosaic_path):
original = self.get_image(self.original)
mosaic = None
step = min(original.height, original.width) / tile_num
for y in range(0, original.height, step):
mosaic_row = None
print('Building row', y/step, '/', original.height/step)
for x in range(0, original.width, step):
tile = self.get_tile(x, y, step)
tile = tile.resize(float(tile_size) / float(min(tile.width, tile.height)))
tile = tile.crop(0, 0, tile_size, tile_size)
#mosaic.draw_image(tile, x, y)
mosaic_row = tile if not mosaic_row else mosaic_row.join(tile, "horizontal")
mosaic = mosaic_row if not mosaic else mosaic.join(mosaic_row, "vertical")
mosaic.write_to_file(mosaic_path)
I have also tried creating a mosaic by resizing the original image and then using draw_image like the following but this also crashes.
mosaic = self.get_image(self.original).resize(tile_size)
mosaic.draw_image(tile, x, y)
Finally, I have tried creating the mosaic from new_temp_file and I am having trouble writing to the temp image.
How can I make this mosaic program work?
libvips uses a recursive algorithm to work out which pixels to compute next, so for very long pipelines you can overflow the C stack and get a crash.
The simplest solution would be to use arrayjoin. This is a libvips operator which can join many images in a single call:
http://jcupitt.github.io/libvips/API/current/libvips-conversion.html#vips-arrayjoin
There's an example on the libvips github of using it to join 30,000 images at once:
https://github.com/jcupitt/libvips/issues/471
(though that's using the previous version of the libvips Python binding)
I adapted your program to use arrayjoin, and changed the way it loaded images. I noticed you were also reloading the original image for each output tile, so removing that gave a nice speedup.
#!/usr/bin/python2
from __future__ import print_function
import os
import sys
import pyvips
from os.path import join
from scipy.spatial import cKDTree
class Mosaic(object):
def __init__(self, dir_path, original_path, tile_size=128, tree=None, averages=None):
self.dir_path = dir_path
self.original_path = original_path
self.tile_size = tile_size
self.tree = tree
if averages:
self.averages = averages
else:
self.averages = {}
def avg_rgb(self, image):
m = image.stats()
return tuple(m(4,i)[0] for i in range(1,4))
def build_tree(self):
for root, dirs, files in os.walk(self.dir_path):
print('Loading images from', root, '...')
for file_name in files:
path = join(root, file_name)
try:
# load image as a square image of size tile_size X tile_size
tile = pyvips.Image.thumbnail(path, self.tile_size,
height=self.tile_size,
crop='centre')
# render into memory
tile = tile.copy_memory()
self.averages[self.avg_rgb(tile)] = tile
except pyvips.error.Error:
print('File', path, 'not recognized as an image.')
self.tree = cKDTree(self.averages.keys())
print('Loaded', len(self.averages), 'images.')
def fetch_tree(self, patch):
avg = self.avg_rgb(patch)
index = self.tree.query(avg)[1]
return self.averages[tuple(self.tree.data[index])]
def make_mosaic(self, tile_num, mosaic_path):
mosaic = None
original = pyvips.Image.new_from_file(self.original_path)
step = min(original.height, original.width) / tile_num
tiles_across = original.width / step
tiles_down = original.height / step
tiles = []
for y in range(0, tiles_down):
print('Building row', y, '/', tiles_down)
for x in range(0, tiles_across):
patch = original.crop(x * step, y * step,
min(step, original.width - x * step),
min(step, original.height - y * step))
tile = self.fetch_tree(patch)
tiles.append(tile)
mosaic = pyvips.Image.arrayjoin(tiles, across=tiles_across)
print('writing ', mosaic_path)
mosaic.write_to_file(mosaic_path)
mosaic = Mosaic(sys.argv[1], sys.argv[2])
mosaic.build_tree()
mosaic.make_mosaic(200, sys.argv[3])
I can run it like this:
$ time ./mosaic2.py samples/ k2.jpg x.png
Loading images from samples/ ...
Loaded 228 images.
Building row 0 / 292
...
Building row 291 / 292
writing x.png
real 7m19.333s
user 7m27.322s
sys 0m30.578s
making a 26496 x 37376 pixel image, in this case, and it runs in about 150mb of memory.