中文简易版文档： https://blog.csdn.net/qq_38451119/article/details/82428612
官方文档：https://imgaug.readthedocs.io/en/latest/source/dtype_support.html
github上的官方文档：https://github.com/aleju/imgaug
例子：https://www.jianshu.com/p/989b2f4b246d
1. 注意： 如果图像维度是[h,w,c]，则要用 images_aug = seq.augment_image(images) ，也就是augment_image后面没有s，与之相反的是augment_images(有s)。
You provided a numpy array of shape (1200, 1920, 3) as input to augment_images(), which was interpreted as (N, H, W). The last dimension however has value 1 or 3, which indicates that you provided a single image with shape (H, W, C) instead. If that is the case, you should use augment_image(image) or augment_images([image]), otherwise you will not get the expected augmentations.
"you will not get the expected augmentations." % (images_copy.shape,))
2. per_channel参数含义: bool or float, optional Whether to use the same value for all channels (False) or to sample a new value for each channel (True). If this value is a float ``p``, then for ``p`` percent of all images`per_channel` will be treated as True, otherwise as False.
实例分析：
import numpy as np
import imgaug as ia
import imgaug.augmenters as iaa
from PIL import Image
import cv2
from functools import reduce
from matplotlib.colors import rgb_to_hsv, hsv_to_rgb
sometimes = lambda aug: iaa.Sometimes(0.9, aug)
sometimes_p1 = lambda aug: iaa.Sometimes(0.9, aug)
def load_batch(batch_idx):
# dummy function, implement this
# Return a numpy array of shape (N, height, width, #channels)
# or a list of (height, width, #channels) arrays (may have different image
# sizes).
# Images should be in RGB for colorspace augmentations.
# (cv2.imread() returns BGR!)
# Images should usually be in uint8 with values from 0-255.
# return np.zeros((128, 32, 32, 3), dtype=np.uint8) + (batch_idx % 255)
im1 = Image.open('./timg.jpeg')
im1= np.array(im1)
return im1
def train_on_images(images,idx):
# dummy function, implement this
#images=cv2.cvtColor(images, cv2.COLOR_RGB2BGR)
cv2.imwrite('./images/'+str(idx)+'_messigray.jpg',images)
pass
seq = iaa.Sequential([
iaa.Fliplr(0.5), # horizontal flips
iaa.Flipud(0.5), # horizontal flips
iaa.Crop(percent=(0, 0.1)), # random crops
# Small gaussian blur with random sigma between 0 and 0.5.
# But we only blur about 50% of all images.
iaa.Sometimes(0.5,
iaa.GaussianBlur(sigma=(0, 0.5))
),
# Strengthen or weaken the contrast in each image.
iaa.ContrastNormalization((0.5, 2.0)),
# Add gaussian noise.
# For 50% of all images, we sample the noise once per pixel.
# For the other 50% of all images, we sample the noise per pixel AND
# channel. This can change the color (not only brightness) of the
# pixels.
iaa.AdditiveGaussianNoise(loc=0, scale=(0,0.1*255), per_channel=0.5),
sometimes_p1(iaa.SaltAndPepper(p=0.02,per_channel=0.5)),
# Make some images brighter and some darker.
# In 20% of all cases, we sample the multiplier once per channel,
# which can end up changing the color of the images.
iaa.Multiply((0.8, 1.2), per_channel=0.2),
# Apply affine transformations to each image.
# Scale/zoom them, translate/move them, rotate them and shear them.
sometimes(iaa.Affine(
scale={"x": (0.9, 1.1), "y": (0.9, 1.1)}, # scale images to 80-120% of their size, individually per axis
translate_percent={"x": (-0.05, 0.05), "y": (-0.05, 0.05)}, # translate by -20 to +20 percent (per axis)
rotate=(-5, 5), # rotate by -45 to +45 degrees
shear=(-5, 5), # shear by -16 to +16 degrees
order=[0, 1], # use nearest neighbour or bilinear interpolation (fast)
cval=(0, 255), # if mode is constant, use a cval between 0 and 255
#mode=ia.ALL # use any of scikit-image's warping modes (see 2nd image from the top for examples)
)),
], random_order=True) # apply augmenters in random order
for batch_idx in range(20):
images = load_batch(batch_idx)
img_height=images.shape[0]
img_width=images.shape[1]
cv2.imwrite('./images/'+'messigray.jpg',images)
images_aug = seq.augment_image(images) # done by the library
train_on_images(images_aug,batch_idx)
同时更改heatmap的方法：
import numpy as np
import imgaug as ia
import imgaug.augmenters as iaa
from PIL import Image
import cv2
from functools import reduce
from matplotlib.colors import rgb_to_hsv, hsv_to_rgb
sometimes = lambda aug: iaa.Sometimes(0.9, aug)
sometimes_p1 = lambda aug: iaa.Sometimes(0.9, aug)
def load_image(batch_idx):
# dummy function, implement this
# Return a numpy array of shape (N, height, width, #channels)
# or a list of (height, width, #channels) arrays (may have different image
# sizes).
# Images should be in RGB for colorspace augmentations.
# (cv2.imread() returns BGR!)
# Images should usually be in uint8 with values from 0-255.
# return np.zeros((128, 32, 32, 3), dtype=np.uint8) + (batch_idx % 255)
im1 = Image.open('./a1_235_1_R.jpg').convert('L')
##print(np.array(im1).shape)
im1= np.array(im1)[np.newaxis, :, :, np.newaxis]
return im1
def load_heatmap(batch_idx):
# dummy function, implement this
# Return a numpy array of shape (N, height, width, #channels)
# or a list of (height, width, #channels) arrays (may have different image
# sizes).
# Images should be in RGB for colorspace augmentations.
# (cv2.imread() returns BGR!)
# Images should usually be in uint8 with values from 0-255.
# return np.zeros((128, 32, 32, 3), dtype=np.uint8) + (batch_idx % 255)
im1 = Image.open('./a1_235_1_R_1.png').convert('L')
im1= np.array(im1)[np.newaxis, :, :, np.newaxis].astype(np.float32)
return im1
def train_on_images(images,idx,i):
# dummy function, implement this
#images=cv2.cvtColor(images, cv2.COLOR_RGB2BGR)
images=images.squeeze(3)
images=images.squeeze(0)
cv2.imwrite('./images/'+str(idx)+"_"+str(i)+'_messigray.jpg',images.astype(np.uint8))
pass
def train_on_heatmaps(images,idx,i):
# dummy function, implement this
#images=cv2.cvtColor(images, cv2.COLOR_RGB2BGR)
images=images.squeeze(3)
images=images.squeeze(0)
images=images*255
cv2.imwrite('./images/'+str(idx)+"_"+str(i)+'_messigray.jpg',images.astype(np.uint8))
pass
seq = iaa.Sequential([
iaa.Fliplr(0.5), # horizontal flips
iaa.Flipud(0.5), # horizontal flips
#iaa.Crop(percent=(0, 0.1)), # random crops
# Small gaussian blur with random sigma between 0 and 0.5.
# But we only blur about 50% of all images.
iaa.Sometimes(0.5,
iaa.GaussianBlur(sigma=(0, 0.2))
),
# Strengthen or weaken the contrast in each image.
iaa.ContrastNormalization((0.6, 1.4)),
# Add gaussian noise.
# For 50% of all images, we sample the noise once per pixel.
# For the other 50% of all images, we sample the noise per pixel AND
# channel. This can change the color (not only brightness) of the
# pixels.
iaa.AdditiveGaussianNoise(loc=0, scale=(0,0.02*255), per_channel=0.5),
sometimes_p1(iaa.SaltAndPepper(p=0.01,per_channel=0.5)),
# Make some images brighter and some darker.
# In 20% of all cases, we sample the multiplier once per channel,
# which can end up changing the color of the images.
iaa.Multiply((0.8, 1.2), per_channel=0.2),
# Apply affine transformations to each image.
# Scale/zoom them, translate/move them, rotate them and shear them.
sometimes(iaa.Affine(
#scale={"x": (0.9, 1.1), "y": (0.9, 1.1)}, # scale images to 80-120% of their size, individually per axis
#translate_percent={"x": (-0.05, 0.05), "y": (-0.05, 0.05)}, # translate by -20 to +20 percent (per axis)
rotate=(-3, 3), # rotate by -45 to +45 degrees
shear=(-3, 3), # shear by -16 to +16 degrees
order=[0, 1], # use nearest neighbour or bilinear interpolation (fast)
cval=(0, 255), # if mode is constant, use a cval between 0 and 255
#mode=ia.ALL # use any of scikit-image's warping modes (see 2nd image from the top for examples)
)),
], random_order=True) # apply augmenters in random order
for batch_idx in range(20):
images = load_image(batch_idx)
heatmaps = load_heatmap(batch_idx)
img_height=images.shape[0]
img_width=images.shape[1]
cv2.imwrite('./images/'+'messigray.jpg',images)
seq_det = seq.to_deterministic() # call this for each batch again, NOT only once at the start
images_aug, heatmaps_aug = seq(images=images,heatmaps=heatmaps) # done by the library
#images_aug = seq(images=images) # done by the library
train_on_images(images_aug,batch_idx, 0)
train_on_heatmaps(heatmaps_aug,batch_idx, 1)
