Feature amplification capsule network for complex images

Abstract

Availability of massive amounts of data is a key contributing factor that influences the performance of deep learning models. Convolutional Neural Networks for instance, require large amounts of data in different variations to enable them generalize well to viewpoints. However, in health and other application domains, data generation and processing tasks are time-consuming and requires annotation by experts. Capsule Network (CapsNet) have been proposed to curtail the limitations of Convolutional Neural Networks (CNNs). Due to the problem of crowding, capsule Networks perform badly on complex and real-life images such as CIFAR 10 and some medical images. In this study, a variant of a capsule network with a new algorithm referred to as the amplifier and a new squash function termed exponential squash is proposed. The amplifier is implemented in the encoder network to improve the texture of the images and has the ability to assign low relevance to irrelevant features and high relevance to vital features. The exponential squash function reduces the coupling strength of unrelated capsules in the lower and upper capsule layer. The proposed algorithm was evaluated on four datasets; CIFAR 10, fashion-MNIST, eye disease dataset and ODIR dataset achieving accuracies of 84.56% 93.76%, 89.02% and 87.27% respectively. This work sheds light on the possibility of applying CapsNet on complex real-world tasks. The proposed model can serve as an intelligent tool to aid medical personnel to diagnose eye disease and apply the necessary treatments.