Optimizing Capsule Network Methods with Norm-based Compression for Image Classification Problems Ridho Nur Rohman Wijaya (a*) Budi Setiyono (a), Mahmud Yunus (a)
a) Departement of Mathematics, Faculty of Scientics, Institut Teknologi Sepuluh Nopember Surabaya, Indonesia
*ridhonurrohmanwijaya[at]gmail.com
Abstract
Capsule Network is a powerful deep learning technique for image classification problems. However, the Capsule Network^s ability to handle images with intricate backgrounds is limited due to difficulty extracting features. In addition, forming capsule forces results in longer computation time and increased training parameters. We propose a mathematical technique by introducing a Norm-based Compression to enhance the performance of the Capsule Network method. The Norm-based Compression, constructed using Euclidean norms, aims to reduce dimensions and accelerate the translation equivariance process within capsules. Our proposed Capsule Network demonstrates significant improvements through extensive experiments on MNIST, Fashion MNIST, and Kuzushiji MNIST datasets, achieving up to 2.61% higher accuracy with the computation speed is 2.73 times faster than the original Capsule Network, and the total parameters utilized during training are reduced by 15%. Our research contributes to expanding the understanding of efficient image classification methods by offering new insights to overcome the limitations of feature extraction.
Keywords: Capsule Network, Deep Learning, Euclidean Norm, Feature Extraction
