Data augmentation is a technique used in machine learning to artificially increase the size of a dataset by applying various transformations to the existing data. This helps improve model generalization, robustness, and performance, especially when working with limited labeled data.
Image Augmentation
Rotation - Rotates images by a certain degree to provide variability.
Flip (Horizontal/Vertical) - Flips images horizontally or vertically to diversify the dataset.
Zoom - Randomly zooms into or out of images, altering their scale.
Crop - Randomly crops regions from images, changing their spatial composition.
Brightness and Contrast - Adjusts the brightness and contrast levels of images to simulate different lighting conditions.
Text Augmentation
Synonym Replacement - Replaces words in sentences with their synonyms to introduce linguistic variations.
Random Insertion - Adds random words into sentences to create more diverse text samples.
Random Deletion - Randomly removes words from sentences, simulating missing or incomplete information.
Random Swap - Swaps positions of words in sentences to generate variations in word order.
Audio Augmentation
Pitch Shifting - Alters the pitch of audio samples to introduce variations.
Time Stretching - Adjusts the duration of audio signals, affecting the speed of playback.
Background Noise Addition - Injects background noise into audio data to simulate different environmental conditions.
Video Augmentation
Frame Sampling - Randomly selects frames from video sequences to create diverse samples.
Speed Variation - Adjusts the playback speed of video clips, introducing temporal variability.
Geometric Transformation
Translation - Shifts the position of data points or objects in images.
Scaling - Adjusts the size of objects or images.
Shearing - Distorts the shape of objects or images by tilting.
Conditional Augmentation - Applies augmentation techniques based on specific conditions or criteria, enhancing the diversity of generated samples.
MixUp, CutOut, and CutMix
MixUp - Mixes samples in the dataset by linearly combining pairs of images and their labels.
CutOut - Removes rectangular regions of the images
CutMix - Replaces rectangular regions of one image with those of another, promoting diverse sample generation.
Style Transfer - Transfers artistic styles to images, creating novel samples while preserving the content.
Generative Adversarial Networks (GANs) - Generates synthetic data using GANs to augment the training dataset.
Sequential Augmentation - Applies a sequence of augmentation techniques to diversify the transformations applied to data samples.