Vector-kernel convolutional neural networks

In computer vision, convolutional neural networks (CNNs) obtain extremely striking recognition performance. However, in many CNNs there exists a great deal of parameter redundancy because of matrix kernels. To address this problem, we propose a novel model, namely, vector-kernel convolutional neural network (VeckerNet). In a VeckerNet, each convolutional layer can only use vector kernels of either size k x 1 or 1 x k. Compared to the popular models, e.g., AlexNet, VGG, ResNet and DenseNet, the VeckerNets obtain up to 20.8% relative performance improvement with the parameter reduction by 3 to 97%. Impressively, compared to the ResNets with the same depth, e.g., 44, 56, 101 and 110 layers, the VeckerNets obtain 0.57 to 1.4% relative performance improvement with a decrease of parameters by up to more than two-thirds. The experimental results indicate that the VeckerNets can retain good recognition performance while effectively reducing network parameters. (C) 2018 Elsevier B.V. All rights reserved.