You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
 
 
 

693 lines
20 KiB

# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import copy
import cv2
import numpy as np
import shapely.ops
from shapely.geometry import Polygon, MultiPolygon, GeometryCollection
from sklearn.linear_model import LinearRegression
from skimage import exposure
from joblib import load
from PIL import Image
def normalize(im, mean, std, min_value=[0, 0, 0], max_value=[255, 255, 255]):
# Rescaling (min-max normalization)
range_value = np.asarray(
[1. / (max_value[i] - min_value[i]) for i in range(len(max_value))],
dtype=np.float32)
im = (im - np.asarray(min_value, dtype=np.float32)) * range_value
# Standardization (Z-score Normalization)
im -= mean
im /= std
return im
def permute(im, to_bgr=False):
im = np.swapaxes(im, 1, 2)
im = np.swapaxes(im, 1, 0)
if to_bgr:
im = im[[2, 1, 0], :, :]
return im
def center_crop(im, crop_size=224):
height, width = im.shape[:2]
w_start = (width - crop_size) // 2
h_start = (height - crop_size) // 2
w_end = w_start + crop_size
h_end = h_start + crop_size
im = im[h_start:h_end, w_start:w_end, ...]
return im
def img_flip(im, method=0):
"""
Flip an image.
This function provides 5 flipping methods and can be applied to 2D or 3D numpy arrays.
Args:
im (np.ndarray): Input image.
method (int|string): Flipping method. Must be one of [
0, 1, 2, 3, 4, 'h', 'v', 'hv', 'rt2lb', 'lt2rb',
'dia', 'adia'].
0 or 'h': flip the image in horizontal direction, which is the most frequently
used method;
1 or 'v': flip the image in vertical direction;
2 or 'hv': flip the image in both horizontal diction and vertical direction;
3 or 'rt2lb' or 'dia': flip the image across the diagonal;
4 or 'lt2rb' or 'adia': flip the image across the anti-diagonal.
Returns:
np.ndarray: Flipped image.
Raises:
ValueError: Invalid shape of images.
Examples:
Assume an image is like this:
img:
/ + +
- / *
- * /
We can flip it with following code:
img_h = img_flip(img, 'h')
img_v = img_flip(img, 'v')
img_vh = img_flip(img, 2)
img_rt2lb = img_flip(img, 3)
img_lt2rb = img_flip(img, 4)
Then we get the flipped images:
img_h, flipped in horizontal direction:
+ + \
* \ -
\ * -
img_v, flipped in vertical direction:
- * \
- \ *
\ + +
img_vh, flipped in both horizontal diction and vertical direction:
/ * -
* / -
+ + /
img_rt2lb, mirrored on the diagonal:
/ | |
+ / *
+ * /
img_lt2rb, mirrored on the anti-diagonal:
/ * +
* / +
| | /
"""
if not len(im.shape) >= 2:
raise ValueError("The number of image dimensions is less than 2.")
if method == 0 or method == 'h':
return horizontal_flip(im)
elif method == 1 or method == 'v':
return vertical_flip(im)
elif method == 2 or method == 'hv':
return hv_flip(im)
elif method == 3 or method == 'rt2lb' or method == 'dia':
return rt2lb_flip(im)
elif method == 4 or method == 'lt2rb' or method == 'adia':
return lt2rb_flip(im)
else:
return im
def horizontal_flip(im):
im = im[:, ::-1, ...]
return im
def vertical_flip(im):
im = im[::-1, :, ...]
return im
def hv_flip(im):
im = im[::-1, ::-1, ...]
return im
def rt2lb_flip(im):
axs_list = list(range(len(im.shape)))
axs_list[:2] = [1, 0]
im = im.transpose(axs_list)
return im
def lt2rb_flip(im):
axs_list = list(range(len(im.shape)))
axs_list[:2] = [1, 0]
im = im[::-1, ::-1, ...].transpose(axs_list)
return im
def img_simple_rotate(im, method=0):
"""
Rotate an image.
This function provides 3 rotating methods and can be applied to 2D or 3D numpy arrays.
Args:
im (np.ndarray): Input image.
method (int|string): Rotating method, which must be one of [
0, 1, 2, 90, 180, 270
].
0 or 90 : rotate the image by 90 degrees, clockwise;
1 or 180: rotate the image by 180 degrees, clockwise;
2 or 270: rotate the image by 270 degrees, clockwise.
Returns:
np.ndarray: Rotated image.
Raises:
ValueError: Invalid shape of images.
Examples:
Assume an image is like this:
img:
/ + +
- / *
- * /
We can rotate it with following code:
img_r90 = img_simple_rotate(img, 90)
img_r180 = img_simple_rotate(img, 1)
img_r270 = img_simple_rotate(img, 2)
Then we get the following rotated images:
img_r90, rotated by 90°:
| | \
* \ +
\ * +
img_r180, rotated by 180°:
/ * -
* / -
+ + /
img_r270, rotated by 270°:
+ * \
+ \ *
\ | |
"""
if not len(im.shape) >= 2:
raise ValueError("The number of image dimensions is less than 2.")
if method == 0 or method == 90:
return rot_90(im)
elif method == 1 or method == 180:
return rot_180(im)
elif method == 2 or method == 270:
return rot_270(im)
else:
return im
def rot_90(im):
axs_list = list(range(len(im.shape)))
axs_list[:2] = [1, 0]
im = im[::-1, :, ...].transpose(axs_list)
return im
def rot_180(im):
im = im[::-1, ::-1, ...]
return im
def rot_270(im):
axs_list = list(range(len(im.shape)))
axs_list[:2] = [1, 0]
im = im[:, ::-1, ...].transpose(axs_list)
return im
def rgb2bgr(im):
return im[:, :, ::-1]
def is_poly(poly):
assert isinstance(poly, (list, dict)), \
"Invalid poly type: {}".format(type(poly))
return isinstance(poly, list)
def horizontal_flip_poly(poly, width):
flipped_poly = np.array(poly)
flipped_poly[0::2] = width - np.array(poly[0::2])
return flipped_poly.tolist()
def horizontal_flip_rle(rle, height, width):
import pycocotools.mask as mask_util
if 'counts' in rle and type(rle['counts']) == list:
rle = mask_util.frPyObjects(rle, height, width)
mask = mask_util.decode(rle)
mask = mask[:, ::-1]
rle = mask_util.encode(np.array(mask, order='F', dtype=np.uint8))
return rle
def vertical_flip_poly(poly, height):
flipped_poly = np.array(poly)
flipped_poly[1::2] = height - np.array(poly[1::2])
return flipped_poly.tolist()
def vertical_flip_rle(rle, height, width):
import pycocotools.mask as mask_util
if 'counts' in rle and type(rle['counts']) == list:
rle = mask_util.frPyObjects(rle, height, width)
mask = mask_util.decode(rle)
mask = mask[::-1, :]
rle = mask_util.encode(np.array(mask, order='F', dtype=np.uint8))
return rle
def crop_poly(segm, crop):
xmin, ymin, xmax, ymax = crop
crop_coord = [xmin, ymin, xmin, ymax, xmax, ymax, xmax, ymin]
crop_p = np.array(crop_coord).reshape(4, 2)
crop_p = Polygon(crop_p)
crop_segm = list()
for poly in segm:
poly = np.array(poly).reshape(len(poly) // 2, 2)
polygon = Polygon(poly)
if not polygon.is_valid:
exterior = polygon.exterior
multi_lines = exterior.intersection(exterior)
polygons = shapely.ops.polygonize(multi_lines)
polygon = MultiPolygon(polygons)
multi_polygon = list()
if isinstance(polygon, MultiPolygon):
multi_polygon = copy.deepcopy(polygon)
else:
multi_polygon.append(copy.deepcopy(polygon))
for per_polygon in multi_polygon:
inter = per_polygon.intersection(crop_p)
if not inter:
continue
if isinstance(inter, (MultiPolygon, GeometryCollection)):
for part in inter:
if not isinstance(part, Polygon):
continue
part = np.squeeze(
np.array(part.exterior.coords[:-1]).reshape(1, -1))
part[0::2] -= xmin
part[1::2] -= ymin
crop_segm.append(part.tolist())
elif isinstance(inter, Polygon):
crop_poly = np.squeeze(
np.array(inter.exterior.coords[:-1]).reshape(1, -1))
crop_poly[0::2] -= xmin
crop_poly[1::2] -= ymin
crop_segm.append(crop_poly.tolist())
else:
continue
return crop_segm
def crop_rle(rle, crop, height, width):
import pycocotools.mask as mask_util
if 'counts' in rle and type(rle['counts']) == list:
rle = mask_util.frPyObjects(rle, height, width)
mask = mask_util.decode(rle)
mask = mask[crop[1]:crop[3], crop[0]:crop[2]]
rle = mask_util.encode(np.array(mask, order='F', dtype=np.uint8))
return rle
def expand_poly(poly, x, y):
expanded_poly = np.array(poly)
expanded_poly[0::2] += x
expanded_poly[1::2] += y
return expanded_poly.tolist()
def expand_rle(rle, x, y, height, width, h, w):
import pycocotools.mask as mask_util
if 'counts' in rle and type(rle['counts']) == list:
rle = mask_util.frPyObjects(rle, height, width)
mask = mask_util.decode(rle)
expanded_mask = np.full((h, w), 0).astype(mask.dtype)
expanded_mask[y:y + height, x:x + width] = mask
rle = mask_util.encode(np.array(expanded_mask, order='F', dtype=np.uint8))
return rle
def resize_poly(poly, im_scale_x, im_scale_y):
resized_poly = np.array(poly, dtype=np.float32)
resized_poly[0::2] *= im_scale_x
resized_poly[1::2] *= im_scale_y
return resized_poly.tolist()
def resize_rle(rle, im_h, im_w, im_scale_x, im_scale_y, interp):
import pycocotools.mask as mask_util
if 'counts' in rle and type(rle['counts']) == list:
rle = mask_util.frPyObjects(rle, im_h, im_w)
mask = mask_util.decode(rle)
mask = cv2.resize(
mask, None, None, fx=im_scale_x, fy=im_scale_y, interpolation=interp)
rle = mask_util.encode(np.array(mask, order='F', dtype=np.uint8))
return rle
def to_uint8(im, stretch=False):
"""
Convert raster data to uint8 type.
Args:
im (np.ndarray): Input raster image.
stretch (bool, optional): Use 2% linear stretch or not. Default is False.
Returns:
np.ndarray: Image data with unit8 type.
"""
# 2% linear stretch
def _two_percent_linear(image, max_out=255, min_out=0):
def _gray_process(gray, maxout=max_out, minout=min_out):
# Get the corresponding gray level at 98% in the histogram.
high_value = np.percentile(gray, 98)
low_value = np.percentile(gray, 2)
truncated_gray = np.clip(gray, a_min=low_value, a_max=high_value)
processed_gray = ((truncated_gray - low_value) / (high_value - low_value)) * \
(maxout - minout)
return np.uint8(processed_gray)
if len(image.shape) == 3:
processes = []
for b in range(image.shape[-1]):
processes.append(_gray_process(image[:, :, b]))
result = np.stack(processes, axis=2)
else: # if len(image.shape) == 2
result = _gray_process(image)
return np.uint8(result)
# Simple image standardization
def _sample_norm(image):
stretches = []
if len(image.shape) == 3:
for b in range(image.shape[-1]):
stretched = exposure.equalize_hist(image[:, :, b])
stretched /= float(np.max(stretched))
stretches.append(stretched)
stretched_img = np.stack(stretches, axis=2)
else: # if len(image.shape) == 2
stretched_img = exposure.equalize_hist(image)
return np.uint8(stretched_img * 255)
dtype = im.dtype.name
if dtype != "uint8":
im = _sample_norm(im)
if stretch:
im = _two_percent_linear(im)
return im
def to_intensity(im):
"""
Calculate the intensity of SAR data.
Args:
im (np.ndarray): SAR image.
Returns:
np.ndarray: Intensity image.
"""
if len(im.shape) != 2:
raise ValueError("`len(im.shape) must be 2.")
# If the type is complex, this is SAR data.
if isinstance(type(im[0, 0]), complex):
im = abs(im)
return im
def select_bands(im, band_list=[1, 2, 3]):
"""
Select bands of a multi-band image.
Args:
im (np.ndarray): Input image.
band_list (list, optional): Bands to select (band index start from 1).
Defaults to [1, 2, 3].
Returns:
np.ndarray: Image with selected bands.
"""
if len(im.shape) == 2: # Image has only one channel
return im
if not isinstance(band_list, list) or len(band_list) == 0:
raise TypeError("band_list must be non empty list.")
total_band = im.shape[-1]
result = []
for band in band_list:
band = int(band - 1)
if band < 0 or band >= total_band:
raise ValueError("The element in band_list must > 1 and <= {}.".
format(str(total_band)))
result.append(im[:, :, band])
ima = np.stack(result, axis=-1)
return ima
def dehaze(im, gamma=False):
"""
Perform single image haze removal using dark channel prior.
Args:
im (np.ndarray): Input image.
gamma (bool, optional): Use gamma correction or not. Defaults to False.
Returns:
np.ndarray: Output dehazed image.
"""
def _guided_filter(I, p, r, eps):
m_I = cv2.boxFilter(I, -1, (r, r))
m_p = cv2.boxFilter(p, -1, (r, r))
m_Ip = cv2.boxFilter(I * p, -1, (r, r))
cov_Ip = m_Ip - m_I * m_p
m_II = cv2.boxFilter(I * I, -1, (r, r))
var_I = m_II - m_I * m_I
a = cov_Ip / (var_I + eps)
b = m_p - a * m_I
m_a = cv2.boxFilter(a, -1, (r, r))
m_b = cv2.boxFilter(b, -1, (r, r))
return m_a * I + m_b
def _dehaze(im, r, w, maxatmo_mask, eps):
# im is a RGB image and the value ranges in [0, 1].
atmo_mask = np.min(im, 2)
dark_channel = cv2.erode(atmo_mask, np.ones((15, 15)))
atmo_mask = _guided_filter(atmo_mask, dark_channel, r, eps)
bins = 2000
ht = np.histogram(atmo_mask, bins)
d = np.cumsum(ht[0]) / float(atmo_mask.size)
for lmax in range(bins - 1, 0, -1):
if d[lmax] <= 0.999:
break
atmo_illum = np.mean(im, 2)[atmo_mask >= ht[1][lmax]].max()
atmo_mask = np.minimum(atmo_mask * w, maxatmo_mask)
return atmo_mask, atmo_illum
if np.max(im) > 1:
im = im / 255.
result = np.zeros(im.shape)
mask_img, atmo_illum = _dehaze(
im, r=81, w=0.95, maxatmo_mask=0.80, eps=1e-8)
for k in range(3):
result[:, :, k] = (im[:, :, k] - mask_img) / (1 - mask_img / atmo_illum)
result = np.clip(result, 0, 1)
if gamma:
result = result**(np.log(0.5) / np.log(result.mean()))
return (result * 255).astype("uint8")
def match_histograms(im, ref):
"""
Match the cumulative histogram of one image to another.
Args:
im (np.ndarray): Input image.
ref (np.ndarray): Reference image to match histogram of. `ref` must have
the same number of channels as `im`.
Returns:
np.ndarray: Transformed input image.
Raises:
ValueError: When the number of channels of `ref` differs from that of im`.
"""
# TODO: Check the data types of the inputs to see if they are supported by skimage
return exposure.match_histograms(
im, ref, channel_axis=-1 if im.ndim > 2 else None)
def match_by_regression(im, ref, pif_loc=None):
"""
Match the brightness values of two images using a linear regression method.
Args:
im (np.ndarray): Input image.
ref (np.ndarray): Reference image to match. `ref` must have the same shape
as `im`.
pif_loc (tuple|None, optional): Spatial locations where pseudo-invariant
features (PIFs) are obtained. If `pif_loc` is set to None, all pixels in
the image will be used as training samples for the regression model. In
other cases, `pif_loc` should be a tuple of np.ndarrays. Default: None.
Returns:
np.ndarray: Transformed input image.
Raises:
ValueError: When the shape of `ref` differs from that of `im`.
"""
def _linear_regress(im, ref, loc):
regressor = LinearRegression()
if loc is not None:
x, y = im[loc], ref[loc]
else:
x, y = im, ref
x, y = x.reshape(-1, 1), y.ravel()
regressor.fit(x, y)
matched = regressor.predict(im.reshape(-1, 1))
return matched.reshape(im.shape)
if im.shape != ref.shape:
raise ValueError("Image and Reference must have the same shape!")
if im.ndim > 2:
# Multiple channels
matched = np.empty(im.shape, dtype=im.dtype)
for ch in range(im.shape[-1]):
matched[..., ch] = _linear_regress(im[..., ch], ref[..., ch],
pif_loc)
else:
# Single channel
matched = _linear_regress(im, ref, pif_loc).astype(im.dtype)
return matched
def match_lf_components(im, ref, lf_ratio=0.01):
"""
Match the low-frequency components of two images.
Args:
im (np.ndarray): Input image.
ref (np.ndarray): Reference image to match. `ref` must have the same shape
as `im`.
lf_ratio (float, optional): Proportion of frequence components that should
be recognized as low-frequency components in the frequency domain.
Default: 0.01.
Returns:
np.ndarray: Transformed input image.
Raises:
ValueError: When the shape of `ref` differs from that of `im`.
"""
def _replace_lf(im, ref, lf_ratio):
h, w = im.shape
h_lf, w_lf = int(h // 2 * lf_ratio), int(w // 2 * lf_ratio)
freq_im = np.fft.fft2(im)
freq_ref = np.fft.fft2(ref)
if h_lf > 0:
freq_im[:h_lf] = freq_ref[:h_lf]
freq_im[-h_lf:] = freq_ref[-h_lf:]
if w_lf > 0:
freq_im[:, :w_lf] = freq_ref[:, :w_lf]
freq_im[:, -w_lf:] = freq_ref[:, -w_lf:]
recon_im = np.fft.ifft2(freq_im)
recon_im = np.abs(recon_im)
return recon_im
if im.shape != ref.shape:
raise ValueError("Image and Reference must have the same shape!")
if im.ndim > 2:
# Multiple channels
matched = np.empty(im.shape, dtype=im.dtype)
for ch in range(im.shape[-1]):
matched[..., ch] = _replace_lf(im[..., ch], ref[..., ch], lf_ratio)
else:
# Single channel
matched = _replace_lf(im, ref, lf_ratio).astype(im.dtype)
return matched
def inv_pca(im, joblib_path):
"""
Perform inverse PCA transformation.
Args:
im (np.ndarray): Input image after performing PCA.
joblib_path (str): Path of *.joblib file that stores PCA information.
Returns:
np.ndarray: Reconstructed input image.
"""
pca = load(joblib_path)
H, W, C = im.shape
n_im = np.reshape(im, (-1, C))
r_im = pca.inverse_transform(n_im)
r_im = np.reshape(r_im, (H, W, -1))
return r_im
def decode_seg_mask(mask_path):
"""
Decode a segmentation mask image.
Args:
mask_path (str): Path of the mask image to decode.
Returns:
np.ndarray: Decoded mask image.
"""
mask = np.asarray(Image.open(mask_path))
mask = mask.astype('int64')
return mask
def calc_hr_shape(lr_shape, sr_factor):
return tuple(int(s * sr_factor) for s in lr_shape)