augmentation.md

Data augmentation transforms

This page lists data augmentation transforms used in DAN.

Individual augmentation transforms

Elastic Transform

	Elastic Transform
Description	This transformation applies local distortions that rotate characters locally.
Comments	The impact of this transformation is mostly visible on documents, not so much on lines. Results are comparable to the original DAN implementation.
Documentation	See the albumentations documentation
Examples
CPU time (seconds/10 images)	0.44 (3013x128 pixels) / 0.86 (1116x581 pixels)

PieceWise Affine

!!! warning This transform is temporarily removed from the pipeline until this issue is fixed.

	PieceWise Affine
Description	This transformation also applies local distortions but with a larger grid than ElasticTransform.
Comments	This transformation is very slow. It is a new transform that was not in the original DAN implementation.
Documentation	See the albumentations documentation
Examples
CPU time (seconds/10 images)	2.92 (3013x128 pixels) / 3.76 (1116x581 pixels)

Dilation Erosion

	Dilation & Erosion
Description	This transformation makes the pen stroke thicker or thinner.
Comments	The RandomDilationErosion class randomly selects a kernel size and applies a dilation or an erosion to the image. It relies on opencv and is similar to the original DAN implementation.
Documentation	See the opencv documentation
Examples
CPU time (seconds/10 images)	0.02 (3013x128 pixels) / 0.03 (1116x581 pixels)

Sharpen

	Sharpen
Description	This transformation makes the image sharper.
Comments	Similar to the original DAN implementation.
Documentation	See the albumentations documentation
Examples
CPU time (seconds/10 images)	0.02 (3013x128 pixels) / 0.04 (1116x581 pixels)

Color Jittering

	Color Jittering
Description	This transformation alters the colors of the image.
Comments	Similar to the original DAN implementation.
Documentation	See the albumentations documentation
Examples
CPU time (seconds/10 images)	0.03 (3013x128 pixels) / 0.04 (1116x581 pixels)

Gaussian Noise

	Gaussian Noise
Description	This transformation adds Gaussian noise to the image.
Comments	The noise from the original DAN implementation is more uniform.
Documentation	See the albumentations documentation
Examples
CPU time (seconds/10 images)	0.29 (3013x128 pixels) / 0.53 (1116x581 pixels)

Gaussian Blur

	Gaussian Blur
Description	This transformation blurs the image.
Comments	Similar to the original DAN implementation.
Documentation	See the albumentations documentation
Examples
CPU time (seconds/10 images)	0.01 (3013x128 pixels) / 0.02 (1116x581 pixels)

Random Perspective

	Random Perspective
Description	This transformation changes the perspective from which the photo is taken.
Comments	Similar to the original DAN implementation.
Documentation	See the albumentations documentation
Examples
CPU time (seconds/10 images)	0.05 (3013x128 pixels) / 0.05 (1116x581 pixels)

Shearing (x-axis)

	Shearing (x-axis)
Description	This transformation changes the slant of the text on the image.
Comments	New transform that was not in the original DAN implementation.
Documentation	See the albumentations documentation
Examples
CPU time (seconds/10 images)	0.05 (3013x128 pixels) / 0.04 (1116x581 pixels)

Coarse Dropout

	Coarse Dropout
Description	This transformation adds dropout on the image, turning small patches into black pixels.
Comments	It is a new transform that was not in the original DAN implementation.
Documentation	See the albumentations documentation
Examples
CPU time (seconds/10 images)	0.02 (3013x128 pixels) / 0.02 (1116x581 pixels)

Random Scale

	RandomScale
Description	This transformation downscales the image from a random factor.
Comments	The original DAN implementation reimplemented it as DPIAdjusting.
Documentation	See the albumentations documentation
Examples

To Gray

	ToGray
Description	This transformation transforms an RGB image into grayscale.
Comments	It is a new transform that was not in the original DAN implementation.
Documentation	See the albumentations documentation
Examples
CPU time (seconds/10 images)	0.02 (3013x128 pixels) / 0.02 (1116x581 pixels)

Full augmentation pipeline

• Data augmentation is applied with a probability of 0.9.
• In this case, two transformations are randomly selected to be applied.
• Reproducibility is possible by setting random.seed and np.random.seed (already done in dan/ocr/document/train.py)
• Examples with new pipeline: