| With the combination of computer vision algorithms and road image processing,it has become possible for artificial intelligence to assist the detection and maintenance of road surface diseases.Training a crack detection model with good performance in practical applications usually requires a large-scale and highly diverse data set,but the acquisition of crack images is a very time-consuming and labor-intensive project,and it is difficult to meet the data requirements of current deep learning network models.In order to solve the problems mentioned,this paper proposes a crack image data set augmentation method based on Generative Adversarial Networks,and conducts in-depth research on it.Firstly,pre-process the acquired crack image to improve the image quality.Aiming at the Gaussian noise and salt and pepper noise existing in the crack image,Gaussian filtering and median filtering are used to denoise,respectively.And through the histogram equalization and gamma transformation,enhance the crack feature recognition,making the crack feature easier to learn by the Convolutional Neural Network.Secondly,Deep Convolutional Generative Adversarial Networks(DCGAN)is used to virtually augment the road crack image data set.In view of the difficulties in training the network model,such as easy to collapse,non-convergence,instability and poor results in generating crack images,an improved method for systematically encoding the real crack image in the sample data using the coding part of Variational Auto-Encoder(VAE)is proposed,and the hidden variable value obtained by the encoder is used as the generator of the DCGAN model input section.According to the test results,the hyperparameters in the model are optimized to achieve the expansion of the crack image data set.Finally,a pavement crack classification detection model based on Faster R-CNN is used to evaluate the reliability of the generated crack image quality.This part mainly evaluates the quality of the generated crack image in the two aspects of the crack image data set before and after augmentation and the crack image data set augmented by the current conventional method.The experiment result shows that the crack image data set augmented by the generative adversarial networks and the crack image data dataset augmented by conventional methods can improve the training effect.But the former augmentation data set has better training effect on the detection model,stronger generalization ability and higher accuracy under the same test set.It is proved that the crack image data set virtual augmentation method proposed in this paper not only can solve the problem of insufficient crack image data set effectively,but also can improve the recognition efficiency and detection accuracy of pavement cracks.The average detection accuracy of cracks reaches 90.32%.At the same time,it provides the possibility of intelligentization of pavement crack detection and reduction of pavement maintenance costs. |
