Dense Blocks-Incorporated Capsule Networks for Image Classification on Complex Datasets

Capsule network (CapsNet) has received a special concern recently owing to its great capability of guaranteeing the model robustness against the affine transformation and adversarial images. However, most CapsNets perform poor while costing huge at image classification on complex datasets, due probably to the weak feature extraction, the gradient problems and many intermediate variables employed in dynamic routing. To address this issue, a novel CapsNet called dense CapsNet (DenseCap) is proposed. First, dense blocks with short-cuts are stacked before capsule layers to adequately extract abundant essential features from image input while avoiding vanishing gradient. Then, CapsNet ReLU (CapsReLU) function is put forward to squash the capsule such that the sensitivity to the parameter initiation and vanishing gradient are handled. Furthermore, the influence of various dimensions of capsules on the reconstructed images is carefully studied. Finally, empirical results on five benchmark datasets show that the proposed DenseCaps outperform state-of-the-art capsule models.