This is a Cutout [1] / Random Erasing [2] implementation. In particular, it is easily used with ImageDataGenerator in Keras. Please check random_eraser.py for implementation details.
Cutout or Random Erasing is a kind of image augmentation methods for convolutional neural networks (CNN). They are very similar methods and were proposed almost at the same time.
They try to regularize models using training images that are randomly masked with random values.
It is very easy to use if you are using ImageDataGenerator in Keras;
get eraser
function by get_random_eraser()
,
and then pass it to ImageDataGenerator
as preprocessing_function
.
By doing so, all images are randomly erased before standard augmentation
done by ImageDataGenerator.
Please check cifar10_resnet.py, which is imported from official Keras examples.
What I did is adding only two lines:
...
from random_eraser import get_random_eraser # added
...
datagen = ImageDataGenerator(
...
preprocessing_function=get_random_eraser(v_l=0, v_h=1)) # added
Of cause, you can erase a single image using eraser
function.
Please note that eraser
function works in inplace mode;
the input image itself will be modified (therefore, img = eraser(img)
can be replaced by eraser(img)
in the following example).
from random_eraser import get_random_eraser
eraser = get_random_eraser()
# load image to img
img = eraser(img)
Pleae check example.ipynb for complete example.
Parameters are fully configurable as:
get_random_eraser(p=0.5, s_l=0.02, s_h=0.4, r_1=0.3, r_2=1/0.3,
v_l=0, v_h=255, pixel_level=False)
p
: the probability that random erasing is performeds_l
, s_h
: minimum / maximum proportion of erased area against input imager_1
, r_2
: minimum / maximum aspect ratio of erased areav_l
, v_h
: minimum / maximum value for erased areapixel_level
: pixel-level randomization for erased areaThe original cifar10_resnet.py
result (w/o cutout / random erasing):
Test loss: 0.539187009859
Test accuracy: 0.9077
With cutout / random erasing:
Test loss: 0.445597583055
Test accuracy: 0.9182
With cutout / random erasing (pixel-level):
Test loss: 0.446407950497
Test accuracy: 0.9213
[1] T. DeVries and G. W. Taylor, "Improved Regularization of Convolutional Neural Networks with Cutout," in arXiv:1708.04552, 2017.
[2] Z. Zhong, L. Zheng, G. Kang, S. Li, and Y. Yang, "Random Erasing Data Augmentation," in arXiv:1708.04896, 2017.