Computer Vision News - September 2019

We Tried for You 16 by Amnon Geifman Data augmentation refers to the process of generating new data from existing one. It is well known that the generalization gap, i.e the difference in performance between the test and the train data, depends on the number of examples in the training set. As the size of the training set increases, the performance is similar on both seen and unseen data. This month we will review a few of the most common techniques to enlarge data sets using several transformations. We will use Keras ImageDataGenerator to create several popular augmentations and explain the rationality behind them. The goal in data augmentation is to generate new examples without changing the class label of the data. A good example of that is shifting. The human eye can identify an object no matter where the object is located on the image. Mathematically two images with the same object in a different location are different; however, we want our network to recognize the object regardless of its location, just like humans do. The rationale behind data augmentation is to show the network enough examples of a shifted object, and in this way to develop invariance of the network for shifting. Of course, shifting is just one example and we will see other transformations as well. Next, we will see how exactly we can create such shifted images without changing the class label and without hurting the naturalness of the image. Remember, we want to take an image and shift the main object to a different direction. In Keras we perform such transformation via the keras. preprocessing.image.ImageDataGenerator class. "Data augmentation is the common practice when training a neural network. It enlarges the data set and enables to boost the generalization capabilities of the network. Keras data generator enables to augment data on the fly without taking care of all the technical details of the transformations." "Training the network on the full data set may produce even better results"