The recognition performance of convolutional neural networks has surpassed that of humans in many computer vision areas. However, there is a large number of parameter redundancy in deep neural networks, especially the weights of the convolutional kernels. In this work, we propose a simple Diagonal-kernel, in which a standard square kernel is replaced by a diagonal kernel and an anti-diagonal kernel. Diagonal-kernels with fewer parameters can have similar or larger local receptive fields than square kernels. The performance of the Diagonal-kernel is firstly evaluated on two benchmark image classification datasets, CIFAR, and ImageNet. The experimental results indicate that the Diagonal-kernel can effectively reduce parameters and computational cost while maintaining high accuracy. Furthermore, compared with Vector-kernel, Diagonal-kernel has larger local receptive fields and is more efficient. Then, we test the Diagonal-kernel for fine-grained image and imbalanced image dataset. The results show that Diagonal-kernel has larger accuracy loss for fine-grained than the coarse-grain image, but the loss is tolerable. The imbalanced data does not influence the performance of the Diagonal-kernel. The proposed Diagonal-kernel is mainly for traditional convolution but not for depthwise convolution because the number of weights for deep convolution is very small.