Integrating Circle Kernels into Convolutional Neural Networks

The square kernel is a standard unit for contemporary Convolutional Neural Networks (CNNs), as it fits well on the tensor computation for the convolution operation. However, the receptive field in the human visual system is actually isotropic like a circle. Motivated by this observation, we propose using circle kernels with isotropic receptive fields for the convolution, and our training takes approximately equivalent amount of calculation when compared with the corresponding CNN with square kernels. Our preliminary experiments demonstrate the rationality of circle kernels. We then propose a kernel boosting strategy that integrates the circle kernels with square kernels for the training and inference, and we further let the kernel size/radius be learnable during the training. Note that we reparameterize the circle kernels or integrated kernels before the inference, thus taking no extra computation as well as the number of parameter overhead for the testing. Extensive experiments on several standard datasets, ImageNet, CIFAR-10 and CIFAR-100, using the circle kernels or integrated kernels on typical existing CNNs, show that our approach exhibits highly competitive performance. Specifically, on ImageNet with standard data augmentation, our approach dramatically boosts the performance of MobileNetV3-Small by 5.20% top-1 accuracy and 3.39% top-5 accuracy, and boosts the performance of MobileNetV3-Large by 2.16% top-1 accuracy and 1.18% top-5 accuracy.