深t度学习中，数据增强是一种不会改变网络模型的正则化方法，通过对数据的增强可提升模型最终的训练效果。 数据增强的两种主要方式： 常用的增强技术： 翻转：对图像进行水平或者垂直翻转。 旋转：对图像给定一个角度进行顺时针或者逆时针旋转。 裁剪：在裁剪的过程中，对图像的一部分进行随机采样 亮度：增加或者减少图像的亮度 缩放：缩放图像可以向外缩放或向内缩放。向外缩放时，图像大小增大，向内缩放时图像大小减小 添加噪声：可以给现存的照片添加高斯噪声

目前主流的图像增强库主要有以下：

skimage opencv imgaug Albumentations Augmentor Keras(ImageDataGenerator class) SOLT

1.使用skimage进行数据增强

scikit-image是一个使用numpy数组的开放源码Python包。它为研究、教育和工业应用实现算法和工具。即使对于那些刚刚接触Python生态系统的人来说，它也是一个相当简单和直接的库。 首先导入必要的库和数据

import os import glob import numpy as np import scipy as sp import pandas as pd #skimage from skimage.io import imshow,imread,imsave from skimage.transform import rotate,AffineTransform,warp,rescale,resize,downscale_local_mean from skimage import color,data from skimage.exposure import adjust_gamma from skimage.util import random_noise #OpenCV-Python import cv2 #imgaug import imageio import imgaug as ia import imgaug.augmenters as iaa #Albumentations import albumentations as A #Augmentor !pip install augmentor import Augmentor #Keras from keras.preprocessing.image import ImageDataGenerator,array_to_img,img_to_array,load_img #SOLT !pip install solt import solt import solt.transforms as slt #visualisation import matplotlib.pyplot as plt import matplotlib.image as mpimg %matplotlib inline import seaborn as sns from IPython.display import HTML,Image #source: https://www.kaggle.com/jpmiller/nfl-punt-analytics/edit # set additional display options for report pd.set_option("display.max_columns',100) th_props=[('font-size','13px'),('background-color','white'),('color','#666666')] td_props = [('font-size', '15px'), ('background-color', 'white')] styles = [dict(selector="td", props=td_props), dict(selector="th", props=th_props)] #warnings import warnings warnings.filterwarnings("ignore") #Helper function to display the images in a grid # Source: https://stackoverflow.com/questions/42040747/more-idiomatic-way-to-display-images-in-a-grid-with-numpy which was pointed by # this excellent article: https://towardsdatascience.com/data-augmentation-for-deep-learning-4fe21d1a4eb9 def gallery(array, ncols=3): ''' Function to arange images into a grid. INPUT: array - numpy array containing images ncols - number of columns in resulting imahe grid OUTPUT: result - reshaped array into a grid with given number of columns ''' nindex, height, width, intensity = array.shape nrows = nindex//ncols assert nindex == nrows*ncols result = (array.reshape(nrows, ncols, height, width, intensity) .swapaxes(1,2) .reshape(height*nrows, width*ncols, intensity)) return result #Defining data path Image_Data_Path = "../input/plant-pathology-2020-fgvc7/images/" train_data = pd.read_csv("../input/plant-pathology-2020-fgvc7/train.csv") test_data = pd.read_csv("../input/plant-pathology-2020-fgvc7/test.csv") # Loading the training images #refer: https://www.kaggle.com/tarunpaparaju/plant-pathology-2020-eda-models def load_image(image_id): file_path=image_id+'.jpg' image=imread(Image_Data_Path+file_path) return image train_image=train_data['image_id'][:50].apply(load_image)

1.1 Flipping with skimage

#Horizontally flipped hflipped_image=np.fliplr(image) #fliplr reverse the order of columns of pixels in matrix #Vertically flipped vflipped_image=np.flipud(image)#flipud reverse the order of rows of pixels in matrix fig,ax=plt.subplots(nrows=1,ncols=3,figsize=(30,16)) ax[0].imshow(image) ax[0].set_title('Original Image',size=30) ax[1].imshow(hflipped_image) ax[1].set_title('Horizontally flipped',size=30) ax[2].imshow(vflipped_image) ax[2].set_title('Vertically flipped',size=30);

1.2 Rotation with skimage

#clockwise rotation rot_clockwise_image=rotate(image,angle=45) #Anticlockwise rotation rot_anticlockwise_image=rotate(image,angle=-45) fig,ax=plt.subplots(nrows=1,ncols,figsize=(30,16)) ax[0].imshow(image) ax[0].set_title('Original Image',size=30) ax[1].imshow(rot_clockwise_image) ax[1].set_title('+45 degree Rotation',size=30) ax[2].imshow(rot_anticlockwise_image) ax[2].set_title('-45 degree rotation',size=30);

1.3 Cropping with skimage

import random import pylab as pl def randRange(a,b): return pl.rand()*(b-a)+a def randomCrop(im): margin=1/3.5 start=[int(randRange(0,im.shape[0]*margin)), int(randRange(0,im.shape[1]*margin))] end=[int(randRange(im.shape[0]*(1-margin),im.shape[0])),int(randRange(im.shape[1]*(1-margin),im.shape[1]))] cropped_image=(im[start[0]:end[0],start[1]:end[1]]) return cropped_image fig,ax=plt.subplots(nrows=1,ncols=2,figsize=(20,12)) ax[0].imshow(image) ax[0].set_title('Original Image',size=20) ax[1].imshow(randomCrop(image)) ax[1].set_title('Cropped',size=20)

1.4 Brightness Manipulation

image_bright=adjust_gamma(image,gamma=0.5,gain=1) image_dark=adjust_gamma(image,gamma=2,gain=1) fig,ax=plt.subplots(nrow=1,ncols=3,figsize=(20,12)) ax[0].imshow(image) ax[0].set_title('Original Image',size=20) ax[1].imshow(image_bright) ax[1].set_title('Brightened Image',size=20) ax[2].imshow(image_dark) ax[2].set_title('Darkened Image',size=20)

1.5 Scaling with skimage

image_resize=resize(image,(image.shape[0]//2,image.shape[1]//2),anti_aliasing=True) #image_downscaled=downscale_local_mean(image,(4,3)) fig,ax=plt.subplots(nrows=1,ncols=2,figsize=(30,16)) ax[0].imshow(image) ax[0].set_title('Original Image',size=20) ax[1].imshow(image_resized) ax[1].set_title('Resized image',size=20)

1.6 Noise Addition with skimage

noisy_image=random_noise(image) fig,ax=plt.subplots(nrows=1,ncols=2,figsize=(30,16)) ax[0].imshow(image) ax[0].set_title('Original Image',size=20) ax[1].imshow(noisy_image) ax[1].set_title('Image after adding noise',size=20)

2.Data Augmentation using OpenCV-Python

OpenCV本质上是开源计算机视觉库的缩写。虽然它是用优化的C/C++编写的，但它有针对Python和Java以及c++的接口。您可以把它看作是OpenCV的c++实现的python包装器。OpenCV-Python不仅速度快(因为后台由C/ c++编写的代码组成)，而且易于编码和部署(因为前台有Python包装器)。这使得它成为执行计算密集型程序的一个很好的选择。

#selecting a sample image image13=train_images[13] imshow(image13) print(image13.shape) plt.axis('off')

2.1 Flipping with opencv

The image is flipped according to the value of flipCode as follows:

flipcode = 0: flip vertically flipcode > 0: flip horizontally flipcode < 0: flip vertically and horizontally

#vertical flip img_flip_ud=cv2.flip(image13,0) plt.imshow(img_flip_ud) #horizontal filp img_flip_lr=cv2.flip(image13,1) plt.imshow(img_flip_lr)

fig,ax=plt.subplots(nrows=1,ncols=2,figsize=(20,12)) ax[0].imshow(img_flip_ud) ax[0].set_title('vertical flip',size=20) ax[1].imshow(img_filp_lr) ax[1].set_title('horizontal flip',size=20)

2.2 Rotation with opencv

OpenCV函数中旋转图像的函数是cv2.rotate() 可以在旋转代码中指定以下三个常量

cv2.ROTATE_90_CLOCKWISE cv2.ROTATE_90_COUNTERCLOCKWISE cv2.ROTATE_180 img_rotate_90_clockwise=cv2.rotate(image13,cv.ROTATE_90_CLOCKWISE) img_rotate_90_counterclockwise=cv2.rotate(image13,cv2.ROTATE_90_COUNTERCLOCKWISE) img_rotate_180=cv2.rotate(image13,cv2.ROTATE_180) fig,ax=plt.subplots(nrows=1,ncols=3,figsize=(20,12)) ax[0].imshow(img_rotate_90_clockwise) ax[0].set_title('90 degrees clockwise',size=20) ax[1].imshow(img_rotate_90_counterclockwise) ax[1].set_title('90 degrees anticlockwise',size=20) ax[2].imshow(img_rotate_180) ax[2].set_title('180 degree rotation',size=20)

2.3 Scaling with opencv

Scaling 是重塑图片的尺寸

#RESIZE def resize_image(image,w,h): resized_image=image=cv2.resize(image,(w,h)) return resize_image imshow(resize_image(image13,500,500))

2.4 Brightness Manipulation with opencv

def add_light(image,gamma): invGamma=1.0/gamma table=np.array([((i/255.0)**invGamma)*255 for i in np.arange(0,256)]).astype('uint8') image=cv2.LUT(image,table) return image imshow(add_light(image13,2))

2.5 Cropping with opencv

#crop def crop_image(image,y1,y2,x1,x2): image=image[y1:y2,x1:x2] return image imshow(crop_image(image13,200,800,250,1500)

2.6 Gaussian Blur with opencv

Gaussian滤波器是低通滤波器，能移除被减少的高频部分

def gaussian_blur(image,blur): image=cv2.GaussianBlur(image,(5,5),blur) return image imshow(gaussian_blur(image13,0))

3.Data Augmentation using imgaug

imgaug是一个机器学习实验中用于图像增强的库。它支持广泛的增强技术，允许轻松地以随机顺序结合和执行他们在多个CPU核心上，有一个简单而强大的随机界面，不仅可以增强图像，也能够用于检测关键点/地标，边界盒，热图和分割地图。

# selecting a sample image image2=train_image[25] imshow(image2) print(image2.shape) plt.axis('off')

3.1 Flipping with imgaug

#Horizontally flipped hflip=iaa.Fliplr(p=1.0) hflipped_image2=hflip.augment_image(image2) #Vertically flipped vflip=iaa.Flipud(p=1.0) vflipped_image2=vflip.augment_image(image2) image=image2 fig,ax=plt.subplots(nrows=1,ncols=3,figsize=(30,16)) ax[0].imshow(image) ax[0].set_title('Original Image',size=30) ax[1].imshow(hflipped_image2) ax[1].set_title('Horizontally flipped',size=30) ax[2].imshow(vflipped_image2) ax[2].set_title('Vertically flipped',size=30)

3.2 Rotation with imgaug

#clockwise rotation rot=iaa.Affine(rotate=(-25,25)) rot_clockwise_image2=rot.augment_image(image2) image=image2 fig,ax=plt.subplots(nrows=1,ncols=2,figsize=(30,16)) ax[0].imshow(image) ax[0].set_title('Original Image',size=30) ax[1].imshow(rot_clockwise_image2) ax[1].set_title('Rotated Image',size=30)

3.3 Cropping with imgaug

image=image2 crop=iaa.Crop(precent=(0,0.2)) corp_image=crop.augment_image(image) fig,ax=plt.subplots(nrows=1,ncols=2,figsize=(20,12)) ax[0].imshow(image) ax[0].set_title('Original Image',size=20) ax[1].imshow(randomCrop(corp_image)) ax[1].set_title('Cropped',size=20)

3.4 Brightness Manipulation with imgaug

image=image2 #bright contrast1=iaa.GammaContrast(gamma=0.5) brightened_image=contrast1.augment_image(image) #dark contrast2=iaa.GammaContrast(gamma=2) darkened_image=contrast2.augment_image(image) fig,ax=plt.subplots(nrows=1,ncols=3,figsize=(20,12)) ax[0].imshow(image) ax[0].set_title('Original Image',size=20) ax[1].imshow(brightened_image) ax[1].set_title('Brightened Image',size=20) ax[2].imshow(darkened_image) ax[2].set_title('darkened_image',size=20)

3.5 Scaling with imgaug

image=image2 scale_img=iaa.Affine(scale={'x':(1.5,1.0),'y':(0.5,1.0)}) scale_image=scale_im.augment_image(image) fig,ax = plt.subplots(nrows=1,ncols=2,figsize=(20,12)) ax[0].imshow(image) ax[0].set_title("Original Image", size=20) ax[1].imshow(scale_image) ax[1].set_title("Scaled", size=20)

3.6 Noise addition with imgaug

image=image2 gaussian_noise=iaa.AdditiveGaussianNoise(15,20) noise_image=gaussian_noise.augment_image(image) fig,ax = plt.subplots(nrows=1,ncols=2,figsize=(20,12)) ax[0].imshow(image) ax[0].set_title("Original Image", size=20) ax[1].imshow(noise_image) ax[1].set_title("Gaussian Noise added", size=20)

3.7 Augmentation pipeline

imgaug库提供了一个非常有用的特性，称为Augmentation pipeline。这样的管道是可以按固定或随机顺序应用的一系列步骤。这也提供了对一些图像应用某些转换和对其他图像应用其他转换的灵活性。在下面的例子中，我们在一些图像上应用翻转、锐化、裁剪等转换。模糊和仿射变换有时会被应用，所有这些变换将会被随机地应用。

#Defining a pipeline #The example has been taken from the documentation aug_pipeline=iaa.Sequential([iaa.SomeOf((0,3),[iaa.Fliplr(1.0),# horizontally flip iaa.Flipud(1.0),# veritical flip iaa.Sharpen(alpha=(0,1.0),lightness=(0.75,1.5)), #sharpen images iaa.Crop(percent=(0,0.4)), iaa.Sometimes(0.5,iaa.Affine(rotate=5)), iaa.Sometimes(0.5,iaa.GaussianBlur(sigma=(0,0.5))), iaa.AdditiveGaussianNoise(loc=0,scale=(0.0,0.05*255),per_channel=0.5), ]) ], random_order=True # apply the augmentations in random order) # apply augmentation pipeline to sample image images_aug = np.array([aug_pipeline.augment_image(image2) for _ in range(16)]) # visualize the augmented images plt.figure(figsize=(30,10)) plt.axis('off') plt.imshow(gallery(images_aug, ncols = 4)) plt.title('Augmentation examples')

4.Data Augmentation using Albumentations

Albumentations是一个快速的图像增强库，易于使用包装其他库。它是基于numpy, OpenCV, imgaug从中选出最好的。它是由Kagglers写的，被用于在Kaggle, topcoder, CVPR, MICCAI的许多DL比赛中获得最佳结果。更多信息请点击:https://www.mdpi.com/2078-2489/11/2/125

#initialize augmentations horizontal_filp=A.HorizontalFlip(p=1) rotate=A.ShiftScaleRotate(p=1) gaus_noise=A.GaussNoise()#gaussian noise bright_contrast = A.RandomBrightnessContrast(p=1) # random brightness and contrast gamma = A.RandomGamma(p=1) # random gamma blur = A.Blur() # apply augmentations to images img_flip = horizontal_flip(image = image2) img_gaus = gaus_noise(image = image2) img_rotate = rotate(image = image2) img_bc = bright_contrast(image = image2) img_gamma = gamma(image = image2) img_blur = blur(image = image2) # access the augmented image by 'image' key img_list = [img_flip['image'],img_gaus['image'], img_rotate['image'], img_bc['image'], img_gamma['image'], img_blur['image']] # visualize the augmented images plt.figure(figsize=(10,10)) plt.axis('off') plt.imshow(gallery(np.array(img_list), ncols = 3)) plt.title('Augmentation examples')

5. Data Augmentation using Augmentor

# Passing the path of the image directory p = Augmentor.Pipeline(source_directory="/kaggle/input/plant-pathology-2020-fgvc7/images", output_directory="/kaggle/output") # Defining augmentation parameters and generating 10 samples p.flip_left_right(probability=0.4) p.flip_top_bottom(probability=0.8) p.rotate(probability=0.5, max_left_rotation=5, max_right_rotation=10) p.skew(0.4, 0.5) p.zoom(probability = 0.2, min_factor = 1.1, max_factor = 1.5) p.sample(10)

6. Keras Image Data Generator

Keras库有一个为向图像添加转换而创建的内建类。这个类称为ImageDataGenerator，它会生成一批张量图像数据并进行实时数据扩充.

# selecting a sample image image5 = train_images[15] imshow(image5) print(image5.shape) plt.axis('off')

# Augmentation process datagen = ImageDataGenerator( rotation_range = 40, shear_range = 0.2, zoom_range = 0.2, horizontal_flip = True, brightness_range = (0.5, 1.5)) img_arr = img_to_array(image5) img_arr = img_arr.reshape((1,) + img_arr.shape) i = 0 for batch in datagen.flow( img_arr, batch_size=1, save_to_dir='../output/keras_augmentations', save_prefix='Augmented_image', save_format='jpeg'): i += 1 if i > 20: # create 20 augmented images break # otherwise the generator would loop indefinitely images = os.listdir("../output/keras_augmentations/") images # Let's look at the augmented images aug_images = [] for img_path in glob.glob("../output/keras_augmentations/*.jpeg"): aug_images.append(mpimg.imread(img_path)) plt.figure(figsize=(20,10)) columns = 5 for i, image in enumerate(aug_images): plt.subplot(len(aug_images) / columns + 1, columns, i + 1) plt.imshow(image)

# selecting a sample image image5 = train_images[25] imshow(image5) print(image5.shape) plt.axis('off')

7. SOLT : Streaming over lightweight data transformations

SOLT是一个快速的数据扩充库，支持任意数量的图像、分割掩码、关键点和数据标签。它的后端有OpenCV，因此工作非常快。

h,w,c = image5.shape img = image5[:w] stream = solt.Stream([ slt.Rotate(angle_range=(-90, 90), p=1, padding='r'), slt.Flip(axis=1, p=0.5), slt.Flip(axis=0, p=0.5), slt.Shear(range_x=0.3, range_y=0.8, p=0.5, padding='r'), slt.Scale(range_x=(0.8, 1.3), padding='r', range_y=(0.8, 1.3), same=False, p=0.5), slt.Pad((w, h), 'r'), slt.Crop((w, w), 'r'), slt.CvtColor('rgb2gs', keep_dim=True, p=0.2), slt.HSV((0, 10), (0, 10), (0, 10)), slt.Blur(k_size=7, blur_type='m'), solt.SelectiveStream([ slt.CutOut(40, p=1), slt.CutOut(50, p=1), slt.CutOut(10, p=1), solt.Stream(), solt.Stream(), ], n=3), ], ignore_fast_mode=True) fig = plt.figure(figsize=(16,16)) n_augs = 6 random.seed(42) for i in range(n_augs): img_aug = stream({'image': img}, return_torch=False, ).data[0].squeeze() ax = fig.add_subplot(1,n_augs,i+1) if i == 0: ax.imshow(img) else: ax.imshow(img_aug) ax.set_xticks([]) ax.set_yticks([]) plt.show()