我想使用 Earth Movers Distance 来比较多个图像。
我将 scipy.stats.wasserstein_distance() 与 pyemd.emd_samples() 进行了比较。 据我了解,wasserstein_distance() 采用两个分布,即直方图,而 emd_samples() 采用一维值数组并为您计算直方图。
鉴于这两种方法使用相同的直方图,它们应该提供相同或至少相似的结果。
问题是,这两种方法提供的结果截然不同。但是,如果我在这两种方法中传递图像的扁平化版本,结果将非常相似。
是我这边有误还是其中一个实现有问题?
cat1 = skimage.io.imread("./cat1.jpg", as_grey=True).flatten().astype('float64')
cat2 = skimage.io.imread("./cat2.jpg", as_grey=True).flatten().astype('float64')
shuttle = skimage.io.imread("./shuttle.jpg", as_grey=True).flatten().astype('float64')
emd_s = np.array([[emd_samples(cat1, cat1, bins="fd"), emd_samples(cat1, cat2, bins="fd"), emd_samples(cat1, shuttle, bins="fd")],
[emd_samples(cat2, cat1, bins="fd"), emd_samples(cat2, cat2, bins="fd"), emd_samples(cat2, shuttle, bins="fd")],
[emd_samples(shuttle, cat1, bins="fd"), emd_samples(shuttle, cat2, bins="fd"), emd_samples(shuttle, shuttle, bins="fd")]])
pmf_cat1, bins_cat1 = np.histogram(cat1 , bins="fd")
pmf_cat2, bins_cat2 = np.histogram(cat2 , bins="fd")
pmf_shuttle, bins_shuttle = np.histogram(shuttle , bins="fd")
emd_s2 = np.array([[emd_samples(pmf_cat1, pmf_cat1, bins="fd"), emd_samples(pmf_cat1, pmf_cat2, bins="fd"), emd_samples(pmf_cat1, pmf_shuttle, bins="fd")],
[emd_samples(pmf_cat2, pmf_cat1, bins="fd"), emd_samples(pmf_cat2, pmf_cat2, bins="fd"), emd_samples(pmf_cat2, pmf_shuttle, bins="fd")],
[emd_samples(pmf_shuttle, pmf_cat1, bins="fd"), emd_samples(pmf_shuttle, pmf_cat2, bins="fd"), emd_samples(pmf_shuttle, pmf_shuttle, bins="fd")]])
swd = np.array([[wasserstein_distance(pmf_cat1, pmf_cat1), wasserstein_distance(pmf_cat1, pmf_cat2), wasserstein_distance(pmf_cat1, pmf_shuttle)],
[wasserstein_distance(pmf_cat2, pmf_cat1), wasserstein_distance(pmf_cat2, pmf_cat2), wasserstein_distance(pmf_cat2, pmf_shuttle)],
[wasserstein_distance(pmf_shuttle, pmf_cat1), wasserstein_distance(pmf_shuttle, pmf_cat2), wasserstein_distance(pmf_shuttle, pmf_shuttle)]])
swd2 = np.array([[wasserstein_distance(cat1, cat1), wasserstein_distance(cat1, cat2), wasserstein_distance(cat1, shuttle)],
[wasserstein_distance(cat2, cat1), wasserstein_distance(cat2, cat2), wasserstein_distance(cat2, shuttle)],
[wasserstein_distance(shuttle, cat1), wasserstein_distance(shuttle, cat2), wasserstein_distance(shuttle, shuttle)]])
上面的示例对 emd_s 和 swd2 产生了相似的结果,对 emd_s2 和 swd 也产生了相似的结果,尽管最后一对仍然有很大不同,因为技术上 emd_samples 应该在这种情况下基于直方图制作直方图。
请您参考如下方法:
我遇到了类似的问题,想在这里记下一些事情。
函数
emd_samples和wasserstein_distance都将(经验)分布中观察到的值作为输入,而不是分布本身。emd函数允许您传递分布,但是,您需要提供度量作为附加参数。此外,在将直方图用作(密度)分布时,您需要对其进行归一化。不展平灰度图像意味着您比较仅适用于
pyemd的 2D-histgramms。
示例用法:
import numpy as np
import skimage
import os
from pyemd import emd, emd_samples
from scipy.stats import wasserstein_distance
# get some test images
img1 = skimage.io.imread(os.path.join(skimage.data_dir, 'astronaut.png'))
img2 = skimage.io.imread(os.path.join(skimage.data_dir, 'camera.png'))
img3 = skimage.io.imread(os.path.join(skimage.data_dir, 'horse.png'))
# flatten them
images = [img.ravel() for img in [img1, img2, img3]]
# compute EMD using values
emd_samples(images[0], images[1]) # 25.57794401220945
wasserstein_distance(images[0], images[1]) # 25.76187896728515
# compute EMD using distributions
N_BINS = 256
hists = [np.histogram(img, N_BINS, density=True)[0].astype(np.float64) for img in images]
mgrid = np.meshgrid(np.arange(N_BINS), np.arange(N_BINS))
metric = np.abs(mgrid[0] - mgrid[1]).astype(np.float64)
emd(hists[0], hists[1], metric) # 25.862491463680065
