Useful real-life images meant for computer vision applications appear in complex forms with varied backgrounds. Convolutional neural networks with their invariance outperforms capsule networks (CapsNets) on such images. Recent research to improve the accuracy of CapsNets on complex images such as CIFAR10 involves the increase in depth of the networks. However, smaller architectures with efficient feature extractors can perform comparatively well with additional advantages of having less computational cost, reduced risk of overfitting, increased suitability for deployment on devices with smaller memories, and enhanced feasibility to perform distributed training across servers and in the cloud. This paper, therefore, proposes a shallow CapsNet based on dynamic routing with a custom squashing function and normalizer to gain all of these advantages. The model is evaluated on three publicly available datasets; fashion-MNIST, CIFAR-10, and tomato datasets. Experimental results show that the proposed model outperforms the original CapsNet by generating fewer parameters and performing comparably well in terms of accuracy with some state-of-the-art multi-lane and ensemble capsule network models.

适用于计算机视觉应用的有用的现实生活图像以各种背景的复杂形式出现。具有不变性的卷积神经网络在此类图像上优于胶囊网络（CapsNets）。最近为提高CapsNets在CIFAR10等复杂图像上的准确性而进行的研究涉及网络深度的增加。但是，具有高效特征提取器的较小体系结构可以具有相对较好的性能，并具有以下优点：计算成本较低，过拟合的风险降低，部署在具有较小内存的设备上的适用性增加以及跨服务器和在云中执行分布式培训的可行性更高。因此，本文提出了一种基于动态路由的浅层CapsNet，它具有自定义的压缩功能和规范化器，以获取所有这些优点。该模型在三个公开可用的数据集上进行了评估。fashion-MNIST，CIFAR-10和Tomato数据集。实验结果表明，与一些最新的多车道和整体胶囊网络模型相比，该模型产生的参数更少并且在准确性方面具有可比性，从而优于原始CapsNet。