augmentation_techniques.md

Overview of Augmentation techniques

We present the different techniques that can be applied to augment a dataset of images. In order to generate the samples presented in this section, it is necessary to load some libraries.

from clodsa.techniques.techniqueFactory import createTechnique
import cv2
import numpy as np

img = cv2.imread("images/cat1.jpg")

Average blurring

Smoothes the image using an average filter.

t = createTechnique("average_blurring", {"kernel" : 5})
img1 = t.apply(img)
cv2.imwrite("average_blurring.jpg",np.hstack([img,img1]))

Bilateral blurring

Applies bilateral blurring to the image.

t = createTechnique("bilateral_blurring", {"diameter" : 11, "sigmaColor": 21, "sigmaSpace":7})
img1 = t.apply(img)
cv2.imwrite("bilateral_blurring.jpg",np.hstack([img,img1]))

Blurring

Blurs an image using the normalized box filter.

t = createTechnique("blurring", {"ksize" : 5})
img1 = t.apply(img)
cv2.imwrite("blurring.jpg",np.hstack([img,img1]))

Change to HSV

Changes the color space from RGB to HSV.

t = createTechnique("change_to_hsv",{})
img1 = t.apply(img)
cv2.imwrite("change_to_hsv.jpg",np.hstack([img,img1]))

Change to LAB

Changes the color space from RGB to LAB.

t = createTechnique("change_to_lab",{})
img1 = t.apply(img)
cv2.imwrite("change_to_lab.jpg",np.hstack([img,img1]))

Crop

Crops pixels at the sides of the image.

t = createTechnique("crop",{"percentage":0.8,"startFrom": "TOPLEFT"})
img1 = t.apply(img)
cv2.imwrite("crop.jpg",img1)

Dropout

Sets some pixels in the image to zero.

t = createTechnique("dropout",{"percentage":0.05})
img1 = t.apply(img)
cv2.imwrite("dropout.jpg",np.hstack([img,img1]))

Elastic deformation

Applies elastic deformation as explained in the paper: P. Simard, D. Steinkraus, and J. C. Platt. Best practices for convolutional neural networks applied to visual document analysis. Proceedings of the 12th International Conference on Document Analysis and Recognition (ICDAR'03) vol. 2, pp. 958--964. IEEE Computer Society. 2003.

t = createTechnique("elastic",{"alpha":5,"sigma":0.05})
img1 = t.apply(img)
cv2.imwrite("elastic.jpg",np.hstack([img,img1]))

Equalize histogram

Applies histogram equalization to the image.

t = createTechnique("equalize_histogram",{})
img1 = t.apply(img)
cv2.imwrite("equalize.jpg",np.hstack([img,img1]))

Flip

Flips the image horizontally, vertically or both.

t = createTechnique("flip",{"flip":0})
t1 = createTechnique("flip",{"flip":1})
t2 = createTechnique("flip",{"flip":-1})
img1 = t.apply(img)
img2 = t1.apply(img)
img3 = t2.apply(img)
cv2.imwrite("flip.jpg",np.hstack([img,img1,img2,img3]))

Gamma correction

Applies gamma correction to the image.

t = createTechnique("gamma",{"gamma":1.5})
img1 = t.apply(img)
cv2.imwrite("gamma.jpg",np.hstack([img,img1]))

Gaussian blurring

Blurs an image using a Gaussian filter.

t = createTechnique("gaussian_blur", {"kernel" : 5})
img1 = t.apply(img)
cv2.imwrite("gaussian_blurring.jpg",np.hstack([img,img1]))

Gaussian noise

Adds Gaussian noise to the image.

t = createTechnique("gaussian_noise", {"mean" : 0,"sigma":10})
img1 = t.apply(img)
cv2.imwrite("gaussian_noise.jpg",np.hstack([img,img1]))

Invert

Inverts all values in images, i.e. sets a pixel from value v to 255-v

t = createTechnique("invert",{})
img1 = t.apply(img)
cv2.imwrite("invert.jpg",np.hstack([img,img1]))

Median blurring

Blurs an image using the median filter.

t = createTechnique("median_blur", {"kernel" : 5})
img1 = t.apply(img)
cv2.imwrite("median_blurring.jpg",np.hstack([img,img1]))

None

This augmentation technique does not change the image.

t = createTechnique("none",{})
img1 = t.apply(img)
cv2.imwrite("none.jpg",np.hstack([img,img1]))

Raise blue channel

Raises the values in the blue channel.

t = createTechnique("raise_blue", {"power" : 0.9})
img1 = t.apply(img)
cv2.imwrite("raise_blue.jpg",np.hstack([img,img1]))

Raise green channel

Raises the values in the green channel.

t = createTechnique("raise_green", {"power" : 0.9})
img1 = t.apply(img)
cv2.imwrite("raise_green.jpg",np.hstack([img,img1]))

Raise hue

Raises the hue value.

t = createTechnique("raise_hue", {"power" : 0.9})
img1 = t.apply(img)
cv2.imwrite("raise_hue.jpg",np.hstack([img,img1]))

Raise red channel

Raises the value in the red channel.

t = createTechnique("raise_red", {"power" : 0.9})
img1 = t.apply(img)
cv2.imwrite("raise_red.jpg",np.hstack([img,img1]))

Raise saturation

Raises the saturation.

t = createTechnique("raise_saturation", {"power" : 0.9})
img1 = t.apply(img)
cv2.imwrite("raise_saturation.jpg",np.hstack([img,img1]))

Raise value

Raise the value of pixels.

t = createTechnique("raise_value", {"power" : 0.9})
img1 = t.apply(img)
cv2.imwrite("raise_value.jpg",np.hstack([img,img1]))

Resize

Resizes the image.

t = createTechnique("resize", {"percentage" : 0.9,"method":"INTER_NEAREST"})
img1 = t.apply(img)
cv2.imwrite("resize.jpg",img1)

Rotate

Rotates the image.

t = createTechnique("rotate", {"angle" : 90})
img1 = t.apply(img)
cv2.imwrite("rotate.jpg",img1)

Salt and Pepper

Adds salt and pepper noise to the image.

t = createTechnique("salt_and_pepper", {"low" : 0,"up":25})
img1 = t.apply(img)
cv2.imwrite("salt_and_pepper.jpg",np.hstack([img,img1]))

Sharpen

Sharpens the image.

t = createTechnique("sharpen", {})
img1 = t.apply(img)
cv2.imwrite("sharpen.jpg",np.hstack([img,img1]))

Shift channel

Shifts the channels of the image.

t = createTechnique("shift_channel", {"shift":0.2})
img1 = t.apply(img)
cv2.imwrite("sharpen.jpg",np.hstack([img,img1]))

Shearing

Shears the image.

t = createTechnique("shearing", {"a":0.5})
img1 = t.apply(img)
cv2.imwrite("shearing.jpg",np.hstack([img,img1]))

Translation

Translates the image.

t = createTechnique("translation", {"x":10,"y":10})
img1 = t.apply(img)
cv2.imwrite("translation.jpg",np.hstack([img,img1]))