A computational ghost imaging method based on convolutional neural network is proposed to solve the imaging quality and the speed of reconstructed images under low sampling condition. Firstly, a convolutional neural network is trained by using a set of training images which are reconstructed by the correlation calculation method and corresponding lossless images. Then, the test set images reconstructed by the correlation calculation are used as the input layer of the convolutional neural network to learn the sensing model and predict the corresponding images. Finally, the images reconstructed by the convolutional neural network are compared with the images reconstructed by computational ghost imaging and compressed sensing algorithm, respectively. The experimental results show that the proposed method can restore the measured object with high quality when the sampling rate is 0.08, and the image quality is higher than other methods. Meanwhile, it takes about 0.06 s without sacrificing image quality when the method is used to reconstruct the single image, which greatly improves the speed of image reconstruction. The effectiveness of our method is also verified by numerical simulation and optical experiments, which is of great significance for engineering applications.