据我所知，ROI由输入图像的connected components给出。每个连接的组件都是一个ROI。
你可以在不迭代ROI的情况下解决这个问题，如果有很多ROI的话，这会变得非常昂贵。但是，这段代码确实显式地遍历了像素，这在Python中很慢。也许，迭代ROI并使用编译函数来处理所有像素会产生更快的Python代码，即使实际上要做更多的工作。但是用编译语言编写，这个算法是你能写的最有效的算法。

import numpy as np
import skimage

image = skimage.io.imread("/Users/cris/Downloads/G4IXb.png")
n_channels = image.shape[2]

# Find binary image separating background and objects
mask = np.any(image > 0, axis=2)

# Label this mask image, each ROI gets a unique label
lab, n_labs = skimage.measure.label(mask, return_num=True)

# Create arrays where we'll collect the information
sum = np.zeros((n_labs + 1, n_channels))
number = np.zeros(n_labs + 1)

# Loop over all pixels.
for ii in range(image.shape[0]):
    for jj in range(image.shape[1]):
        # For each pixel, look at the label, and add information to
        # the arrays for that label.
        index = lab[ii, jj]
        sum[index, :] += image[ii, jj, :]
        number[index] += 1

# We're done. The mean color is the sum divided by the number of pixels.
mean_colors = sum / number[:, None]

字符串
让我们看看平均颜色是什么样子的：

# Paint an image with the mean colors
mean_colors = np.round(mean_colors).astype(np.uint8)
out = np.empty_like(image)
for ii in range(image.shape[0]):
    for jj in range(image.shape[1]):
        out[ii, jj] = mean_colors[lab[ii, jj]]

import matplotlib.pyplot as plt
plt.subplot(1, 2, 1)
plt.imshow(image)
plt.subplot(1, 2, 2)
plt.imshow(out)
plt.show()

型

的数据
当然，由于我们使用skimage，我们也可以使用skimage.measure.regionprops，并获得intensity_mean属性：

props = skimage.measure.regionprops(lab, image)
cols = np.zeros((n_labs + 1, n_channels))
for p in props:
    cols[p.label, :] = p.intensity_mean

型