Augmentation strategies

While deep learning techniques have been extremely successful in computer vision tasks like image recognition, segmentation, or object detection, the underlying algorithms are typically extremely data-intensive: they require large amounts of data to avoid overfitting. However, not all domains of interest satisfy that requirement, which is where data augmentation comes in. This is the name for a group of image processing techniques that create modified versions of images, thus enhancing the size and quality of training datasets, leading to better performance of deep learning models. The augmented data will typically represent a more comprehensive set of possible data points, thereby minimizing the distance between the training and validation set, as well as any future test sets.

In this section, we will review some of the more common augmentation techniques, along with choices for their software implementations. The most frequently used transformations include...