如何估计skimage中的仿射变换矩阵？答案

【问题标题】：How to estimate affine transformation matrix in skimage?如何估计skimage中的仿射变换矩阵？
【发布时间】：2020-07-29 17:54:58
【问题描述】：

我需要将三角形图像（右上图，红色）转换到另一个位置（右上图，绿色）。在this example 之后，我正在尝试估计仿射矩阵并将其应用于转换。但结果不对（左下图）。

在下面的代码中，我试图从uv_coords_src（右上图，红色）转换为uv_coords_dst（右上图，绿色）

import numpy as np

from skimage import io
from skimage import draw
from skimage import transform
from skimage import img_as_float
from skimage import data

from matplotlib import pyplot as plt

uv_coords_src = np.array([[239,287], [101,340], [96,196]])
uv_coords_dst = np.array([[253,179], [170,70], [263,46]])

img = img_as_float(data.chelsea())

mask = draw.polygon2mask(image_shape=img.shape, polygon=uv_coords_src)
mask_out = draw.polygon2mask(image_shape=img.shape, polygon=uv_coords_dst)
masked_image = img * mask
masked_image_out = img * (mask + mask_out)

tform3 = transform.AffineTransform()
tform3.estimate(uv_coords_src, uv_coords_dst)
warped = transform.warp(masked_image, tform3, output_shape=img.shape)

fig, ((ax1, ax2), (ax3, ax4)) = plt.subplots(nrows=2, ncols=2)
ax1.imshow(img, cmap='gray')
ax2.imshow(masked_image_out, cmap='gray')
ax3.imshow(warped, cmap='gray')

plt.show()

请帮我解决它。

【问题讨论】：

标签： transformation scikit-image affinetransform

【解决方案1】：

似乎skimage.transform.warp() 无法正常工作..

我通过scipy.ndimage.affine_transform()进行了改造

import numpy as np

from skimage import io
from skimage import draw
from skimage import transform
from skimage import img_as_float
from skimage import data

import scipy

from matplotlib import pyplot as plt

uv_coords_src = np.array([[239,287], [101,340], [96,196]])
uv_coords_dst = np.array([[253,179], [170,70], [263,46]])

img = img_as_float(data.chelsea())

mask = draw.polygon2mask(image_shape=img.shape, polygon=uv_coords_src)
mask_out = draw.polygon2mask(image_shape=img.shape, polygon=uv_coords_dst)
masked_image = img * mask
masked_image_out = img * (mask + mask_out)

tform3 = transform.AffineTransform()
tform3.estimate(uv_coords_src, uv_coords_dst)
warped0 = scipy.ndimage.affine_transform(masked_image[:, :, 0], np.linalg.inv(tform3.params))
warped1 = scipy.ndimage.affine_transform(masked_image[:, :, 1], np.linalg.inv(tform3.params))
warped2 = scipy.ndimage.affine_transform(masked_image[:, :, 2], np.linalg.inv(tform3.params))

warped = np.dstack([warped0, warped1, warped2])

fig, ((ax1, ax2), (ax3, ax4)) = plt.subplots(nrows=2, ncols=2)
ax1.imshow(img, cmap='gray')
ax2.imshow(masked_image_out, cmap='gray')
ax3.imshow(warped, cmap='gray')

plt.show()

【讨论】：