In the pavement crack segmentation task, the accurate pixel-level labeling required in the fully supervised training of deep neural networks (DNN) is challenging. Although cracks often exhibit low-level image characters in terms of edges, there might be various high-level background information based on the complex pavement conditions. In practice, crack samples containing various semantic backgrounds are scarce. To overcome these problems, we propose a novel method for augmenting the training data for DNN based crack segmentation task. It employs the generative adversarial network (GAN), which utilizes a crack-free image, a crack image, and a corresponding image mask to generate a new crack image. In combination with an auto-encoder, the proposed GAN can be used to train crack segmentation networks. By creating a manual mask, no additional crack images are required to be labeled, and data augmentation and annotation are achieved simultaneously. Our experiments are conducted on two public datasets using five segmentation models of different sizes to verify the effectiveness of the proposed method. Experimental results demonstrate that the proposed method is effective for crack segmentation.