Patch-Based Capsule Network for Complex Images

Abstract

Capsule Networks are neural networks that have the advantage of learning spatial and hierarchical information from data. They can learn and extract knowledge from smaller datasets (unlike other neural network algorithms); however, they perform poorly on complex and low-resolution images due to a problem called “crowding.” Crowding is attributed to the CapsNets’ attempt to read every object in an image (including background objects), resulting in poor performance. Therefore, this paper proposes a patch-based capsule network and a new squash function (power-B) to decompose an input image into smaller parts enabling the model to focus more on the relevant regions of interest. Experimental results show that the proposed model has efficient feature extraction capabilities, reduced computational time, and a fewer trainable number of parameters. The model’s performance is comparable to the state-of-the-art capsule network models by achieving overall recognition accuracies of 94.62%, 75.68%, and 92.82% for fashion-MNIST, CIFAR 10, and polyp datasets, respectively.