In situations where we can’t collect data from other resources, datasets are too small, or the collected data is not well represented, we need to somehow generate data ourselves. This is called data augmentation. Smartly generated data can tackle many problems, including imbalanced datasets, not enough training data and overfitting.

Data augmentation is usually done as part of your input data pipeline that feeds your model while training. Randomly, instead of feeding an original training image, you will instead apply some augmentations to change it. There are many ways to do data augmentation but some examples are:

• Add noise
• Apply geometric transformations
• Swap color channels
• Random color perturbations
• Adjust brightness/contrast/hue

• simply add augmentations that are similar to what the network is having problems generalising to, for example, your model does not work for black and white images; simply add this as a new augmentation