Deep Learning Essentials

Data augmentation

One of the most common tricks for improving the recognition performance is to augment the training data in an intelligent way. There are multiple strategies to achieve this effect:

• Translation and rotation invariance: For the network to learn translation as well as rotation invariances, it is often suggested to augment a training dataset of images with the different perspective transformation of images. For instance, you can take an input image and flip it horizontally and add it to the training dataset. Along with horizontal flips, you can translate them by a few pixels among other possible transformations.
• Scale invariance: One of the limitations of a CNN is its ineffectiveness to recognize objects at different scales. ...