augment-auto

A python image augmentation library based on opencv and numpy. It can be used for augmenting images in both image classification and object detection tasks. Many different techniques of augmentation are supported, which can be clustered into three major types - geometric transformations, photometric transformations and kernel-based transformations. Library has support for images with bounding boxes as well.

Installation

• Install using pip:

pip install augment-auto

• Install by building from scratch:

git clone https://github.com/keshavoct98/image-augmentation
cd image-augmentation
python setup.py install

Features

1. Geometric Features - Image augmentation with geometric transformation of images.

   • crop(img, point1, point2, box = None)

     Returns cropped image. Image is cropped using point1(x1, y1) and point2(x2, y2).

     1. img = numpy.ndarray

        Image to be cropped.

     2. point1 = tuple of int

        initial crop coordinates in the format - (x1, y1).

     3. point2 = tuple of int

        final crop coordinates in the format - (x2, y2).

     4. box = list, default = None

        Coordinates of bounding box in the format - (x1,y1,x2,y2). If bounding box coordinates are passed, new coordinates are calculated and returned along with output image.

   • rotate(img, angle, keep_resolution = True, box = None)

     Returns image rotated at the given angle.

     1. img = numpy.ndarray

        Image to be rotated.

     2. angle = integer or float

        value of angle at which image is to be rotated.

     3. keep_resolution = bool, default = True

        If True, resolution of image remains same after rotation, else resolution is changed.

     4. box = list

        Coordinates of bounding box in the format - (x1,y1,x2,y2). If bounding box coordinates are passed, new coordinates are calculated and returned along with output image.

   • scale(img, fx, fy, keep_resolution = False, box = None)

     Returns scaled image.

     1. img = numpy.ndarray

        Image to be scaled.

     2. fx = integer or float

        scaling value for x-axis.

     3. fy = integer or float

        scaling value for y-axis.

     4. keep_resolution = bool, default = False

        If True, resolution of image remains same after scaling, extra region is cropped out.

     5. box = list

        Coordinates of bounding box in the format - (x1,y1,x2,y2). If bounding box coordinates are passed, new coordinates are calculated and returned along with output image.

   • shear(img, shear_val, axis = 0, box = None)

     Returns sheared image along given axis.

     1. img = numpy.ndarray

        Image to be sheared.

     2. shear_val = integer or float

        shearing magnitude for given axis.

     3. axis = {0,1}, default = 0

        0 for shear along x-axis, 1 for shear along y-axis.

     4. box = list

        Coordinates of bounding box in the format - (x1,y1,x2,y2). If bounding box coordinates are passed, new coordinates are calculated and returned along with output image.

   • translate(img, tx, ty, box = None)

     Returns translated image.

     1. img = numpy.ndarray

        Image to be translated.

     2. tx = integer or float

        translation magnitude along x-axis.

     3. ty = integer or float

        translation magnitude along y-axis.

     4. box = list

        Coordinates of bounding box in the format - (x1,y1,x2,y2). If bounding box coordinates are passed, new coordinates are calculated and returned along with output image.

   # Geometric Transformations
   
   img = cv2.imread('images/3.jpg')
   
   img_new = crop(img, point1 = (100, 100), point2 = (450, 400))
   
   img_new = rotate(img, angle = 15, keep_resolution = True)
   
   img_new = scale(img, fx = 1.5, fy = 1.5, keep_resolution = False)
   
   img_new = shear(img, shear_val = 0.2, axis = 1)
   
   img_new = translate(img, tx = 50, ty = 60)
   

   

   # Geometric Transformations with bounding box
   
   img = cv2.imread('images/0.jpeg')
   bbox = [581, 274, 699, 321]
   
   img_new, bbox_new = crop(img, point1 = (100, 100), point2 = (650, 400), box = bbox)
   
   img_new, bbox_new = rotate(img, angle = 15, keep_resolution = True, box = bbox)
   
   img_new, bbox_new = scale(img, fx = 1.5, fy = 1.3, keep_resolution = False, box = bbox)
   
   img_new, bbox_new = shear(img, shear_val = 0.2, axis = 0, box = bbox)
   
   img_new, bbox_new = translate(img, tx = 50, ty = 160, box = bbox)
   

   

2. Photometric Features - Image augmentation with photometric transformation of images.

   • brightness_contrast(img, alpha = 1.5, beta = 0)

     Returns image with new pixel intensities.

     img_new = img * alpha + beta

     1. img = numpy.ndarray

        Image whose brightness and contrast has to be modified.

     2. alpha = integer or float, non-negative, default = 1.5

        All pixel values of the passed image are multiplied by value of alpha.

     3. beta = integer or float, default = 0

        Vaue of beta is added to all pixel values of the passed image after multiplication of pixel values with value of alpha.

   • colorSpace(img, colorspace = ‘hsv’)

     Returns image converted to the new colorspace. Three types of colorspace are supported - HSV, YCrCb, LAB.

     1. img = numpy.ndarray

        Image whose colorspace has to be converted.

     2. colorspace = {‘hsv’, ‘ycrcb’, ‘lab’}, default = ‘hsv’

        Colorspace to which image is to be converted.

   • addNoise(img, noise_type = ‘gaussian’, mean = 0, var = 0.05, sp_ratio = 0.5, noise_amount = 0.02)

     Returns image with added noise. Three different types of noise are supported - GAUSSIAN, Salt n Pepper, Poisson.

     1. img = numpy.ndarray

        Image to which noise has to be added.

     2. noise_type = {‘gaussian’, ‘salt_pepper’, ‘poisson’}, default = ‘gaussian

        Type of noise to add.

     3. mean = int or float, (required only with noise_type = ‘gaussian’)., default = 0

        Gaussian noise is generated using mean value.

     4. var = int or float, non-negative, (required only with noise_type = ‘gaussian’)., default = 0.05

        Gaussian noise is generated from the standard deviation calculated value of variance provided.

     5. sp_ratio = int or float, range :- 0 <= sp_ratio <= 1, (required only with noise_type = ‘salt_pepper’)., default = 0.5

        Percentage of salt noise and pepper noise. if value passed is equal to 1, only salt noise is present. Similarly if value is 0, only pepper noise is present.

     6. noise_amount = int or float, non-negative, (required only with noise_type = ‘salt_pepper’ or ‘poisson’)., default = 0.02

        magnitude of salt n pepper/poisson noise is calculated using noise_amount.

   # Photometric Transformations
   
   img = cv2.imread('images/1.jpg')
   img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
   
   img_new = brightness_contrast(img, alpha = 1.3, beta = 20)
   img_new = brightness_contrast(img, alpha = 0.7, beta = -10)
   
   img_new = colorSpace(img, colorspace = 'hsv')
   img_new = colorSpace(img, colorspace = 'ycrcb')
   img_new = colorSpace(img, colorspace = 'lab')
   
   img_new = addNoise(img, 'gaussian', mean = 0, var = 0.08)
   img_new = addNoise(img, 'salt_pepper', sp_ratio = 0.5, noise_amount = 0.1)
   img_new = addNoise(img, 'poisson', noise_amount = 0.5)
   

   

3. Kernel-based features

   • blur(img, blur_type = ‘avg’, ksize = (5, 5), median_ksize = 5, gaussian_sigma = 0)

     Returns blurred image. Three different types of blurring are supported - GAUSSIAN, Salt n Pepper, Poisson.

     1. img = numpy.ndarray

        Image to be blurred.

     2. blur_type = {‘avg’, ‘gaussian’, ‘median’}, default = ‘avg’

        Type of blurring to perform.

     3. ksize = tuple of odd positive integers, (required only with blur_type = ‘avg’ or ‘gaussian’)., default = (5, 5)

        kernel size used for average or gaussian blurring.

     4. median_ksize = int, odd positive integer, (required only with blur_type = ‘median’)., default = 5

        kernel size used for median blurring.

     5. gaussian_sigma = int or float, (required only with blur_type = ‘gaussian’)., default = 0

        Standard deviation used to calculate gaussian kernel.

   • randomErase(img, size, box = None)

     A random rectangular region is erased and replaced by mean value of image pixels. Returns modified image.

     1. img = numpy.ndarray

        Image to be modified.

     2. size = tuple of int

        Size of rectangular region to erase.

     3. box = list

        Coordinates of bounding box in the format - (x1,y1,x2,y2). If bounding box coordinates are passed, rectangular region is erased from the bounding box region.

   • randomCropAdd(img, size, box = None)

     A random rectangular region is erased and added to another region of image. Returns modified image.

     1. img = numpy.ndarray

        Image to be modified.

     2. size = tuple of int

        Size of rectangular region to erase and add.

     3. box = list

        Coordinates of bounding box in the format - (x1,y1,x2,y2). If bounding box coordinates are passed, rectangular region is cropped from and added to the bounding box region.

   • sharpen(img)

     Returns sharpened image.

     1. img = numpy.ndarray

        Image to be sharpened.

   # Kernel-based Transformations
   
   img = cv2.imread('images/0.jpeg')
   bbox = [581, 274, 699, 321]
   
   img_new = randomErase(img, size = (100, 100))
   
   img_new = randomCropAdd(img, size = (100, 100))
   
   img_new = sharpen(img)
   
   img_new = randomErase(img, size = (60, 40), box = bbox)
   
   img_new = randomCropAdd(img, size = (60, 40), box = bbox)
   
   img_new = blur(img, 'avg', ksize = (9,9))
   img_new = blur(img, 'gaussian', ksize = (9,9), gaussian_sigma = 0)
   img_new = blur(img, 'median', median_ksize = 11)
   

   

LICENSE

MIT License

Copyright (c) 2020 keshav sharma

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.