A capsule neural network faces significant challenges in achieving high accuracy on complex datasets due to its high computational complexity and limited ability to represent features. To overcome these limitations, this paper proposes a novel framework, Dempster-Shafer Deep Capsule Attention Model (DDCAM), that combines 3D convolution with attention mechanisms to improve feature representation. This model utilizes the Dempster-Shafer Theory to manage uncertainty and optimize feature selection, resulting in more reliable and accurate learning. In this architecture, dense capsule blocks are used to improve gradient flow and mitigate the effects of vanishing gradients. The DDCAM consistently outperforms state-of-the-art capsule networks on a wide range of datasets, especially complex datasets such as MNIST, CIFAR10, SVHN, and Fashion MNIST. DDCAM improves accuracy by 1.29% on CIFAR10 and 0.76% on SVHN compared to existing models, while achieving precision improvements of 1.22% and 1.08% on CIFAR10 and SVHN, respectively. Additionally, experiments conducted on the AffNIST dataset demonstrate the robustness of DDCAM in handling affine transformations, where it achieved significantly higher accuracy compared to existing models. These results illustrate the model’s enhanced ability to capture complex data relationships and improve both accuracy and precision, advancing capsule neural network performance.

Keywords