用pyvips制作巨大的图像马赛克
Making a huge image mosaic with pyvips
我正在尝试使用 pyvips 制作图像马赛克生成器。所以基本上，给定一个图像（以下称为原始图像）创建一个新的、更大的图像，类似于原始图像，除了每个像素（或更现实的像素组）被更小的不同图像块替换。
我被 pyvips 吸引是因为据说它可以处理巨大的图像并且它可以处理图像而不必将它们完全加载到内存中。
但是，我在创建空白马赛克然后填充平铺图像时遇到问题。
在下面的代码中，我尝试将图块逐行连接在一起以创建马赛克，但不幸的是，这段代码占用了我的 RAM，并且总是出现段错误。
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)
我也试过通过调整原始图像的大小然后使用 draw_image 来创建马赛克，如下所示，但这也会崩溃。
mosaic = self.get_image(self.original).resize(tile_size)
mosaic.draw_image(tile, x, y)
最后，我尝试从 new_temp_file 创建马赛克，但在写入临时图像时遇到问题。
我怎样才能让这个马赛克程序工作？
libvips 使用递归算法计算出接下来要计算的像素，因此对于很长的管道，您可能会溢出 C 堆栈并导致崩溃。
最简单的解决方案是使用 arrayjoin。这是一个 libvips 运算符，可以在一次调用中加入许多图像：
http://jcupitt.github.io/libvips/API/current/libvips-conversion.html#vips-arrayjoin
libvips github 上有一个使用它一次连接 30,000 张图像的示例：
https://github.com/jcupitt/libvips/issues/471
（虽然这是使用以前版本的 libvips Python 绑定）
我修改了你的程序以使用 arrayjoin，并更改了它加载图像的方式。我注意到您还为每个输出图块重新加载了原始图像，因此删除它可以实现很好的加速。
#!/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])
我可以运行这样：
$ 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
制作一个 26496 x 37376 像素的图像，在这种情况下，它 运行s 在大约 150mb 的内存中。
我正在尝试使用 pyvips 制作图像马赛克生成器。所以基本上，给定一个图像（以下称为原始图像）创建一个新的、更大的图像，类似于原始图像，除了每个像素（或更现实的像素组）被更小的不同图像块替换。
我被 pyvips 吸引是因为据说它可以处理巨大的图像并且它可以处理图像而不必将它们完全加载到内存中。
但是，我在创建空白马赛克然后填充平铺图像时遇到问题。
在下面的代码中，我尝试将图块逐行连接在一起以创建马赛克，但不幸的是，这段代码占用了我的 RAM，并且总是出现段错误。
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)
我也试过通过调整原始图像的大小然后使用 draw_image 来创建马赛克，如下所示，但这也会崩溃。
mosaic = self.get_image(self.original).resize(tile_size)
mosaic.draw_image(tile, x, y)
最后，我尝试从 new_temp_file 创建马赛克，但在写入临时图像时遇到问题。
我怎样才能让这个马赛克程序工作？
libvips 使用递归算法计算出接下来要计算的像素，因此对于很长的管道，您可能会溢出 C 堆栈并导致崩溃。
最简单的解决方案是使用 arrayjoin。这是一个 libvips 运算符，可以在一次调用中加入许多图像：
http://jcupitt.github.io/libvips/API/current/libvips-conversion.html#vips-arrayjoin
libvips github 上有一个使用它一次连接 30,000 张图像的示例：
https://github.com/jcupitt/libvips/issues/471
（虽然这是使用以前版本的 libvips Python 绑定）
我修改了你的程序以使用 arrayjoin，并更改了它加载图像的方式。我注意到您还为每个输出图块重新加载了原始图像，因此删除它可以实现很好的加速。
#!/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])
我可以运行这样：
$ 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
制作一个 26496 x 37376 像素的图像，在这种情况下，它 运行s 在大约 150mb 的内存中。