def distort_elastic(image, smooth=10.0, scale=100.0, seed=0):
"""
Elastic distortion of images.
Channel axis in RGB images will not be distorted but grayscale or
RGB images are both valid inputs. RGB and grayscale images will be
distorted identically for the same seed.
Simard, et. al, "Best Practices for Convolutional Neural Networks
applied to Visual Document Analysis",
in Proc. of the International Conference on Document Analysis and
Recognition, 2003.
:param ndarayy image: Image of shape [h,w] or [h,w,c]
:param float smooth: Smoothes the distortion.
:param float scale: Scales the distortion.
:param int seed: Seed for random number generator. Ensures that for the
same seed images are distorted identically.
:return: Distorted image with same shape as input image.
:rtype: ndarray
"""
# create random, smoothed displacement field
rnd = np.random.RandomState(int(seed))
h, w = image.shape[:2]
dxy = rnd.rand(2, h, w, 3) * 2 - 1
dxy = gaussian_filter(dxy, smooth, mode="constant")
dxy = dxy / np.linalg.norm(dxy) * scale
dxyz = dxy[0], dxy[1], np.zeros_like(dxy[0])
# create transformation coordinates and deform image
is_color = len(image.shape) == 3
ranges = [np.arange(d) for d in image.shape]
grid = np.meshgrid(*ranges, indexing='ij')
add = lambda v, dv: v + dv if is_color else v + dv[:, :, 0]
idx = [np.reshape(add(v, dv), (-1, 1)) for v, dv in zip(grid, dxyz)]
distorted = map_coordinates(image, idx, order=1, mode='reflect')
return distorted.reshape(image.shape)
评论列表
文章目录
