Figure 4.

Convolutional neural network structure of diploS/HIC. The CNN takes a single image as input and then passes that image to three three convolution layer units each with different filter sizes to capture variation in the image at different physical scales. Each convolution unit consists of two convolution layers followed by a max pooling and a dropout layer. The outputs from the convolutional units are then concatenated and fed to two fully connected dense layers, each themselves followed by dropout. Finally a softmax activation layer is applied to get a categorical classification.