It is a hot issue that speeding up the network layers and decreasing the network parameters in convolutional neural networks (CNNs). In this paper, we propose a novel method, namely, symmetric decomposition of convolution kernels (SDKs). It symmetrically separates k×k convolution kernels into (k×1 and 1×k) or (1×k and k×1) kernels. We conduct the comparison experiments of the network models designed by SDKs on MNIST and CIFAR-10 datasets. Compared with the corresponding CNNs, we obtain good recognition performance, with 1.1×-1.5× speedup and more than 30% reduction of network parameters. The experimental results indicate our method is useful and effective for CNNs in practice, in terms of speedup performance and reduction of parameters.