Combining

The techniques or filters in geometric transformations can be readily combined with most image topics. For example, you can mix horizontal flip, cropping, resizing, padding, rotation, translation, and noise injection for many domains, such as people, landscapes, cityscapes, and others.

In addition, taking landscape as a topic, you can combine many filters in photometric transformations, such as darkening, lightening, color saturation, and contrast. Hue shifting and color casting may not apply to landscape photos. However, advanced photographic transformation filters, such as adding rain or snow to landscape images, are acceptable.

There’s more: you can add random erasing to landscape images. As a result, 1,000 landscape images may increase to 200,000 photos for training. That is the power of image augmentation.

Fun challenge

Here is a thought experiment: should you augment the entire image dataset or only a segment?

Data augmentation can generate hundreds...