Computational ghost imaging uses the second-order coherence of light fields to reconstruct images. In the case of an unknown disturbance (like atmospheric turbulence) to the probe light, the actual light field reaching the object cannot be calculated, and the images will become blurred if they are reconstructed according to the calculated light field without disturbance. In this paper, we proposed a deep learning based image classification-restoration method to suppress the influence of atmospheric turbulence on computational ghost imaging. Specifically, the classification network based on a convolutional neural network classified images according to their blur degree. Then, the images of each class were restored by the restoration network based on a generative adversarial network. Furthermore, we established a compressive-sensing-based computational ghost imaging model including atmospheric turbulence. As a result, the blurred images caused by atmospheric turbulence of different intensities were obtained, and the blurred images were classified and restored by the deep learning method. The simulation results show that the proposed classification-restoration network can effectively improve the image quality of ghost imaging and significantly improve the structural similarity and peak signal-to-noise ratio of the restored images. Besides, this network can generalize different types of targets.