Convolutional neural networks (CNNs) have been successfully applied to a variety of vision tasks, and most of this success comes from research on network structure. However, in spite that convolution is the most important component of CNN, little research has been done about convolution itself, and attempts have been made recently to overcome the limitation of existing convolution. The shape of conventional convolution is restricted, and it is assigned at the stage of configuration of network structure. Usually, extending the size of convolution or stacking more convolution is used for enlarging the receptive field of the network, but this approaches might reduce the efficiency of network due to increases of network depth or the number of parameters. In this dissertation, we describe methodologies for improving the performance of the network through manipulating convolution. The performance of the neural network can be evaluated in terms of accuracy and speed. The earlier part of this paper, we describe the research related to the improvement of accuracy and the studies about improving the speed of the network in the latter part.