MC-CapsNet : Low-level Pooling and High-level Fusion of Multi-columns Capsule Network

Abstract

Capsule network is a new type of neural network structure. The biggest feature of capsule network is “vector in vector out”, which replaces the previous “scalar in scalar out”. The pristine capsule network completely discards the pooling layers and uses a convolutional kernel size of 9 × 9 to increase the perceptual field However, since the original capsule network has only two feature extraction layers, it cannot extract high level abstract features of objects, which are very important for image classification. Moreover, the original capsule network completely discards the use of pooling layers. Based on the above problems, we propose a multi-column capsule network (MC-CapsNet) with low-level pooling and high-level fusion. Unlike the traditional capsule network, MC-CapsNet has the following features: (a)MC-CapsNet consists of two columns of capsule network: the primary column of capsule network extracts high-level impalpable features; the other column of capsule network extracts subordinate shallow features (b)Pooling operation is added to the initial two convolutional layers of the two columns of capsule network (c)Using the high-level theoretical features extracted from the first capsule network to fuse features with the low-level surface layers extracted from the second capsule network to achieve feature complementarily (d)We add the elements of the digital capsule layers in the two-column capsule network to increase the probability of object features being detected by the capsule. The experimental results show that the classification accuracy of MC-CapsNet reaches 87.96% and 94.42% on the CIFAR10 and FashionMNIST datasets.