Result: An Improved Routing based Capsule Network for Hyperspectral Image Classification

Capsule networks have emerged as a solution to the limitations faced by convolutional neural networks. This innovative architecture focuses on encoding features and capturing spatial relationships within images. Instead of employing max pooling, capsule networks introduce a dynamic routing process. The Capsule network is trained to classify each pixel in a hyperspectral image to predefined categories with suitable loss functions and techniques for optimization. By effectively modeling the complex information embodied in hyperspectral data, capsule networks have the potential to improve the accuracy of hyperspectral image classification, making them an invaluable tool for applications such as remote sensing that rely significantly on spectral information. However, the original three-layer capsule network with dynamic routing exhibits subpar performance on intricate datasets like CIFAR-10/100, SVHN and PaviaU HIS dataset, primarily due to the computationally intensive nature of the dynamic routing algorithm. To tackle these challenges, an enhanced capsule network has been proposed, integrating a dense block layer and an "Improved Routing" algorithm. This improved capsule network configuration has undergone testing on CIFAR-10 and SVHN datasets, resulting in notable enhancements such as improved accuracy, reduced loss rates, and decreased time complexity.