This repo is based on Benchmarking Neural Network Robustness to Common Corruptions and Perturbations and the original repository
-
Make a new conda environment
conda create -n ip_env python=3.8
-
Install Wand package for Ubuntu
sudo apt-get install libmagickwand-dev
-
Install python packages from requirements.txt
pip install -r requirements.txt
To use the scripts for augmenting a dataset of images, follow the steps below:
-
Make a JSON for randomization and effects customization. Copy the
/configs/base.json{ "corruption_methods": { "random_erase": false, "color_jitter": true, "gaussian_noise": true, "shot_noise": true, "impulse_noise": true, "speckle_noise": true, "gaussian_blur": true, "glass_blur": true, "defocus_blur": true, "motion_blur": true, "zoom_blur": true, "fog": true, "spatter": true, "earth_bg": false }, "ordered_randomization_layers":{ "background" : {"effects": ["earth_bg"], "p": 0, "level": null}, "color" : {"effects": ["color_jitter"], "p": 0.3, "level": 1}, "noise" : {"effects": ["gaussian_noise", "shot_noise", "impulse_noise", "speckle_noise"], "p": 0.4, "level": 1}, "blur" : {"effects": ["gaussian_blur", "defocus_blur", "motion_blur", "zoom_blur", "spatter"], "p": 0.3, "level": 1}, "erase" : {"effects": ["random_erase"], "p": 0, "level": null} }, "extensions": { "original": "png", "background":"png", "target": "png" }, "opts":{ "earth_augmentation":{ "ann_file": "C:\\Users\\apoca\\Desktop\\data\\Envisat_ICTR\\json_data\\bbox\\train_bbox.json", "earth_bg_images_dir": "C:\\Users\\apoca\\Google Drive ([email protected])\\Envisat\\data\\Himawari-8\\01_June_2020_0000_1300" }, "skip_existing_target_images": false } }-
Save it in the configs dir.
-
To run the augmentations on a set of images use:
python augment_image_files.py --src-dir=data/src/ --dst-dir=data/dst/ --config=configs/{YOURCONFIG}.json
-
-
Gaussian Noise: Gaussian noise is added by perturbing the normalized pixel intensities in the image with pixel intensity drawn from a normal distribution. The severity is specified by the standard deviation of such a normal distribution. The Level-1 standard deviation is set to 0.08 and the Level-2 standard deviation is set to 0.12.
x = np . array ( img ) / 255. corrup_img = np . clip (x + np . random . normal ( size =x. shape , scale = severity ) , 0,1) * 255
-
Shot Noise: Shot noise is added with the pixel intensity drawn from a Poisson distribution. The severity is used to specify the variance of the distribution, from which the the noise for that pixel would be generated. The Level-1 severity factor is set to 60 and the Level-2 severity factor is set to 25.
x = np . array ( img ) / 255. corrup_img = np . clip ( np . random . poisson (x* severity ) / severity , 0, 1) *255
-
Impulse Noise : Impulse noise is added by replacing a proportion of pixels in the image with hot pixels (normalized pixel intensity = 1). The severity specified the proportion of the total image pixels to be replaced. The Level-1 severity factor is set to 0.015 and the Level-2 severity factor is set to 0.06
x = np . array ( img ) / 255. corrup_img = np . clip ( sk . util . random_noise ( x , mode =’s&p ’, amount = c) ,0 , 1) *255
-
Speckle Noise: Speckle noise is added by perturbing the image pixels by an amount obtained by multiplying pixel intensities with a random value drawn from a Gaussian distribution. The severity specifies the standard deviation of the Gaussian distribution. The Level-1 severity factor is set to 0.15 and the Level-2 severity factor is set to 0.2
x = np . array ( img ) / 255. corrup_img =( x + x * np . random . normal ( size =x . shape , scale = severity ) , 0, 1) *255
-
Gaussian Blur: Gaussian Blur is added by using a convolution operation and modifying the pixel intensity value using a Gaussian kernel. The severity specifies the standard deviation of the Gaussian kernel, which is convolved. The Level-1 severity is set to 1 and the Level-2 severity is set to 2
x = np . array ( img ) / 255. corrup_img = skimage . filters . gaussian (x , sigma = severity , multichannel = True ) ,0, 1) * 255
-
Defocus Blur: Defocus Blur is added by constructing a kernel that represents an aliasing disk with Gaussian blur. The image is then convolved with the disk kernel. The severity specifies a 2-tuple with the radius of the aliasing disk and the standard deviation of the gaussian blur used for the disk. The Level-1 severity factor is set to (3,0.1) and the Level-2 severity factor is set to (4,0.5)
x = np . array ( img ) / 255. kernel = AugmentationHelpers . disk ( radius =c [0] , alias_blur = c [1]) channels = [] for d in range (3) : channels . append ( cv2 . filter2D (x [: , :, d] , -1 , kernel )) channels = np . array ( channels ). transpose ((1 , 2, 0) ) # 3 x224x224 -> 224 x224x3 corrup_img = np . clip ( channels , 0, 1) * 255
-
Motion Blur: Motion Blur is added using the Wand Library 2 . The severity specifies a 2-tuple with the radius of the aliasing disk and the standard deviation of the blur used with the disk. The Level-1 severity factor is set to (7,3) and the Level-2 severity factor is set to (15,5)
-
Zoom Blur: Motion Blur is added by adding two overlays of the zooming on an image array. First overlay zooms in by a large zoom factor and the second overlay zooms out by a small zoom factor. The zoomed image pixel values after each zooming operation are determined by spline interpolation. The severity specifies a 2-tuple with the zoom factor for the two operations. The Level-1 severity factor is set to (1.11,0.01) and the Level-2 severity factor is set to (1.16,0.02)
-
Spatter: Spatter is added by simulating liquid droplets. The description of magnitude is not trivial, and the reader is referred to the original code here 3.