Image-based segmentation and quantification of weak interlayers in rock tunnel face via deep learning

In this paper, an advanced integrated pixel-level method based on the deep convolutional neural network (DCNN) approach named DeepLabv3+ is proposed for weak interlayers detection and quantification. Furthermore, a database containing 32,040 images of limestone, dolomite, loess clay, and red clay is established to verify this method. The proposed model is then trained, validated, and tested via feeding multiple weak interlayers. Moreover, robustness and adaptability of the proposed model are evaluated, and the weak interlayers are extracted. Compared with the fully convolutional network (FCN)-based method and traditional image techniques, the proposed model provides higher accuracy in terms of boundary recognition. Besides, it can further detect multiple weak interlayers at the pixel level in practice. The results reveal that the proposed model can efficiently segment damage for rock tunnel faces, eliminate more noises, and consequently provide a much faster running speed. •An image database of weak interlayers was established using an ADPS.•A CNN-based method was developed for quantitative segmentation.•Three segmentation methods were experimented and compared.•Applicability and robustness were tested, employing a complex rock-face database.