我有以下代码,用于根据分割步骤中创建的标签图像中的掩码,将图像的每个值替换为中值(或任何其他函数)值。感觉for循环好像可以向量化。执行此操作的最佳方法是什么?
我考虑过为每个标签构建一个单独的索引数组,但最终没有看到这会有什么帮助。
import numpy as np
from skimage.segmentation import slic
from skimage import data, io
def create_segment_image(original_image, labels_image):
segment_image = np.zeros(original_image.shape, original_image.dtype)
for label in np.unique(labels_image):
segment_image[labels_image==label] = np.median(original_image[labels_image==label], axis=0)
return segment_image
if __name__ == '__main__':
original_image = data.astronaut()
labels_image = slic(original_image, n_segments=1000, max_iter=10, compactness=7, sigma=1, convert2lab=True, enforce_connectivity=True, min_size_factor=0.1, max_size_factor=3, slic_zero=False)
segment_image = create_segment_image(original_image, labels_image)
# io.imsave('images/segment_image.png', segment_image)
请您参考如下方法:
我不知道如何向量化最内层循环。对 median 的调用每次都会计算不同数量的元素,这使得很难将所有调用放入一个数组中。
另一方面,就如何按标签选择元素而言,有一些相当容易实现的成果。您在原始函数中两次找到每个标签的索引,仅计算一次索引数组可减少大约 25% 的运行时间
def create_segment_image_2(original_image, labels_image):
segment_image = np.zeros(original_image.shape, original_image.dtype)
for label in np.unique(labels_image):
inds = np.where(labels_image == label)
segment_image[inds] = np.median(original_image[inds], axis=0)
return segment_image
您可以通过按标签对数组索引进行排序,然后利用该排序来选择图像元素进行中值处理,从而获得更大的改进。用单一排序代替许多搜索可提供大约 20 倍的加速。
def create_segment_image_3(original_image, labels_image):
segment_image = np.zeros(original_image.shape, original_image.dtype)
# sort the indices by their labels
labelinds = np.argsort(labels_image, None)
labels = np.unique(labels_image)
# use the searchsorted to find the indices for each label
rights = np.searchsorted(labels_image.flatten(), labels, side='right', sorter=labelinds)
left = 0
for right in rights:
# choose our block of the image array
inds = labelinds[left:right]
# convert back to a two dimensional index array
inds = [inds // original_image.shape[1], inds % original_image.shape[1]]
segment_image[inds] = np.median(original_image[inds], axis=0)
# update our boundaries
left = right
return segment_image
ipython 中的基准测试
In [54]: %timeit create_segment_image(original_image, labels_image)
2.15 s ± 29.5 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
In [55]: %timeit create_segment_image_2(original_image, labels_image)
1.48 s ± 4.68 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
In [56]: %timeit create_segment_image_3(original_image, labels_image)
121 ms ± 561 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
确认我们的新解决方案提供与旧解决方案相同的结果。
In [57]: np.all(create_segment_image_2(original_image, labels_image) == create_segment_image(original_image, labels_image))
Out[57]: True
In [58]: np.all(create_segment_image_3(original_image, labels_image) == create_segment_image(original_image, labels_image))
Out[58]: True
