Max-Pooled Fast Learning Gabor Capsule Network

Capsule Networks have shown good performance on traditional datasets such as MNIST. However, they are unable to replicate this excellent performance when classifying images with varied backgrounds. This paper further explores this problem by proposing a variant of the Capsule network to investigate the possibility of improving their performance on other image datasets. By substituting global Gabor filter banks followed by Max Pooling in place of Convolutional Neural Network, the initial layers are given the ability to encode texture and spatial relationships leading to improved performance. Max Pooling is used to ensure only dominant features are selected after convolution by Gabor filters resulting in the reduction of the dimensionality of feature vectors. Experiments on four image datasets show that the proposed model has superior performance over the baseline Capsule network achieving very high values for precision, recall, specificity, and accuracy. Also, the proposed model converges faster and has relatively fewer parameters compared to the baseline CapsNet.