Data augmentation-assisted muck image recognition during shield tunnelling

This paper proposed a framework for muck types identification based on data augmentation-assisted image recognition during shield tunnelling. The muck pictures were collected from the shield monitoring system above the conveyor belt. The data augmentation operations were then used to increase the quality of the original images. Furthermore, the Bayesian optimisation algorithm was employed to adjust the parameters of augmenters and highlight the features of the photos. The deep image recognition algorithms (AlexNet and GoogLeNet) were trained and enhanced by the augmentation images, which were used to establish the muck types identification models and assessed by the evaluation indices. Model efficiency was analysed through the performance and time cost of training and validation processes to select the optimal model for muck types identification. Results showed that the performance of identification models could be highly increased by data augmentation with Bayesian optimisation, and the enhanced GoogLeNet performed the highest efficiency for muck types identification.