Feature and spatial relationship coding capsule network

Abstract

A capsule network encodes entity features into a capsule and maps a spatial relationship from the local feature to the overall feature by dynamic routing. This structure allows the capsule network to fully capture feature information but inevitably leads to a lack of spatial relationship guidance, sensitivity to noise features, and easy susceptibility to falling into local optimization. Therefore, we propose a novel capsule network based on feature and spatial relationship coding (FSc-CapsNet). Feature and spatial relationship extractors are introduced to capture features and spatial relationships, respectively. The feature extractor abstracts feature information from bottom to top, while attenuating interference from noise features, and the spatial relationship extractor provides spatial relationship guidance from top to bottom. Then, instead of dynamic routing, a feature and spatial relationship encoder is proposed to find the optimal combination of features and spatial relationships. The encoder abandons the idea of iterative optimization but adds the optimization process to the backpropagation. The experimental results show that, compared with the capsule network and its multiple derivatives, the proposed FSc-CapsNet achieves significantly better performance on both the Fashion-MNIST and CIFAR-10 datasets. In addition, compared with some mainstream deep learning frameworks, FSc-CapsNet performs quite competitively on Fashion-MNIST.