Network Set-Up for Image Classification

Given the basic topology of a deep convolutional neural network for end-to-end learning of image classification, we will discuss details to create such model for a labeling task. Important components of a network design are the number of layers in a convolutional building block, the number of filters in each layer, the perceptive field, the kind of boundary treatment for convolution layers, and the kind of pooling. We will discuss different parameterizations for each of these hyperparameters in view of a classification task as well as strategies to gradually develop a network with a satisfactory performance. We will also take a closer look at the activation function and discuss a number of alternatives to the standard ReLU activation. The second part of this chapter treats the data set-up. Often images come in different sizes and shapes. They need to be resampled, rescaled, and mapped to a common aspect ratio in order to be acceptable as network input. In order to deal with the scarcity of training images in view of the large number of trainable parameters, we introduce basic data augmentation techniques to extend the number of samples by transformation of geometry and illumination of existing training samples.