Convolutional Neural Network gained lots of success in 2012 thanks to AlexNet, a Neural Network capable of excellent performance on the ImageNet dataset. Since then numerous variants of CNNs have been developed, pushing the limits of the architecture for image classification.
At the end of 2017 S. Sabour, N. Frosst and G. Hinton published a paper named “Dynamic Routing Between Capsules” in which a new architecture for object recognition was proposed: the Capsule Network. The idea behind this architecture is to give the classifier a deeper understanding of the objects contained inside the image, encoding the entities that form the object inside vectors of instantiation parameters, representing the pose of the entity.
This is achieved using groups of neurons, named capsules, representing the entities present in the image. Moreover, a new algorithm for the training of layers of capsules is presented. The “Routing- by-Agreement” algorithm is an iterative algorithm which “routes” the activations from one layer of capsules to the next, based on the prediction that each capsule of one layer makes
about the pose of the entity encoded in the capsules of the next layer.
The original paper about Capsule Networks only describes results obtained in small and simple datasets such as MNIST and CIFAR-10. The goal of this thesis is to verify the possible applications of Capsule Networks in place of Convolutional Neural Networks. CNNs are used today in a variety of applications; in this thesis, we will discuss Image Classification, Image Segmentation, and Generative Adversarial Networks. The experiments presented in this thesis concern aspects of Capsule Networks that were never discussed in the original paper, including big input images, different datasets, time of training, and different applications.
This project has been developed in collaboration with AddFor S.p.A. , which has provided significant support and advice.