A Comparative Analysis of a State-of-the-Art CNN versus a Bespoke Capsule Network for Cell Image Classification

Abstract

Despite state of the art performance on object recognition and image classification problems, CNNs are considered to have two significant weaknesses. Firstly, their inability to cater for changes in object orientation, position or lighting. Secondly, their inability to deal with part-whole relationships between objects. Capsule Networks are an enhancement to CNNs to more closely model the viewpoint invariance capability of human vision. The application of Capsule Networks to well known datasets, such as MNIST and NORB, has achieved state of the art performance, while application to other datasets has had mixed results. The application of Capsule Networks to domains such as medical based imaging problems is of significant interest as they have been shown to train accurately on some datasets with limited training data. The contribution of this research is to compare the performance of a Capsule Network to a highly accurate CNN specifically developed for classification of malaria infected and uninfected cell images. It looks at how the accuracy of each model is affected by the volume of available training data, and at how robust each model is to classifying test images subjected to transformation such as rotation, shear and lighting change.

Recommended citation: Liam Murphy, Ruairi O'Reilly (2020). “A Comparative Analysis of a State-of-the-Art CNN versus a Bespoke Capsule Network for Cell Image Classification.” Irish Machine Vision & Image Processing Conference.