Generation method of pavement crack images based on deep convolutional generative adversarial networks

An asphalt pavement crack image generation method was proposed based on deep convolutional generative adversarial network(DCGAN) to improve the quality of a specific pavement image dataset. Firstly, the crack images were captured by a combination of in-vehicle motion camera photography and manual cell phone photography to obtain a more balanced and feature-rich sample small image set. Secondly, original images were denoised by filtering and gamma transformed to enhance the recognition of crack features in the plots, so that a training set of asphalt pavement crack data was created. Thirdly, a deep convolutional generative adversarial neural network model was constructed. The parameters of asphalt pavement crack image generation network were adjusted and its network hyperparameters was optimized to achieve a more realistic generation of pavement crack image dataset. Finally, faster regional convolutional neural network (R-CNN) detection network was used to detect the generated crack images, which can verify the effectiveness of the generated images in the detection network. The results show that the method based on deep convolutional generative adversarial network can generate more realistic crack images. The proposed method can address the problem of insufficient quantity and low quality of datasets under specific conditions more effectively than conventional augmentation methods. Inputting generated virtual images and real pavement images into the detection model can improve the pavement crack detection accuracy. © 2021, Central South University Press. All right reserved.