Research On Shadow Image Data Augmentation For Pavement Detection

With the rapid development of deep learning,pavement detection has begun to change from traditional manual visual detection to automated detection.In order to accurately evaluate the safety of road structure and the performance of road,accurate and rapid pavement damage detection is one of the important conditions.However,due to many complex interferences in the actual pavement detection environment,automatic pavement detection still faces challenges.Among them,the shadows everywhere on the pavement are the most important reason for affecting the automatic road detection.Aiming at the problem of shadow interference in road detection,most of the current research focuses on shadow removal preprocessing of images,but this method not only increases the complexity of automatic road detection,but also puts forward high requirements for the number of datasets.In order to simplify the preprocessing of automatic pavement detection and reduce the training cost of deep learning networks,this paper performs data augmentation on pavement images with shadows to solve the problem of poor crack segmentation in the case of pavement with shadows.Firstly,for the pavement shadow image,due to the mutual interference between cracks and shadows,the generation effect of the original Cycle GAN network is not good.In this paper,we propose a Texture Self-Supervised Cycle GAN(Cycle GAN-TSS)network to improve the shadow generation effect.And it can be used for data augmentation of shadow images with cracks on the pavement.Several images were selected from three public datasets: Crack500,cracktree200 and CFD to create a pavement image dataset,and some images with shadows were manually produced.The above images were sent to Cycle GAN-TSS network for training,and compared with the generated images of the original Cycle GAN under the same circumstances.It is found that the effect of road shadow image generated by the network is better.Then,the effect of the proposed method on crack segmentation under shadow interference is verified.A U-Net network for pavement crack segmentation detection is built,and the model training segmentation results of the augmented dataset and the unaugmented dataset are qualitatively and quantitatively compared.The comparison results show that the segmentation network achieves higher crack recognition accuracy after training on the augmented image dataset.It is proved that the method of augmented dataset can effectively improve the accuracy of crack detection and segmentation under shadow interference.Finally,the method is verified on the actual pavement image dataset collected on campus.The results show that part of the pavement crack images without shadow can be augmented into pavement images with shadow by sending them into the corresponding generation network.The input of these augmented image datasets can effectively improve the accuracy of crack segmentation under shadow interference after model transfer training,which shows that this method has good practical application effect.The method proposed in this paper is to augment the pavement image dataset with shadow and send it to the network for training,and finally improve the accuracy of the detection network for crack segmentation under shadow interference.The research results on the public dataset and the actual collected campus pavement image dataset show that the method is feasible.This study can also provide a new idea for improving the detection accuracy under other interference conditions in future pavement recognition work.