Abstract

By quantitatively measuring the region of the weak interlayer on sites to assess the structure stability of rock tunnel face has formed a significant part under tunneling construction[1,2].However,the scale of images collected from Mengzi-Pingbian highway tunnel(MPHT)in Yunnan,China varies owing to the dif ferent sampling distances between the tunnel face and photographic system.It may lead unsatisfactory segmentation results in overall task.A pixel-level segmentation[3]deep learning model(Figure 1)based on the convolutional neural network(CNN),known as Deep Lab V3+[4,5],was proposed in this paper.The basic architecture of the algorithm was then modified,trained and tested in this paper.Meanwhile,a database containing 3 2 0 4 0 images o f limestone,dolomite,red clay and loess clay were labeled manually.Then,images with multiple sizes were applied into the CNN model to verify the robustness and applicability of the model.

Figure 1　Schematic diagram of the proposed Deep Lab V3+framew ork with a revised encoder-decoder[6,7]and depthwise separable convolution[8]structure

Compared the mean pixel accuracy(MPA)and mean intersection of union(MIo U)[9,10]by testing the multiple image scale(Table 1)from the minimum size(216×230 pixels)to the maximum(3 240×3 450 pixels),the proposed model exhibited a better performance for the small-size images than the large one.In this task,the image size from 216×230 pixels to 864×920 pixels provided higher MPA and MIo U(Figure 2),achieving 88.24% and 78.24%.The image size larger than 1 512×1 380 pixels had a poor performance in term of the boundary segmentation and the noise points.

Table 1　Size statistics of testin g images for deep learning

Figure 2　The MPA and MIoU of different image sizes tested in the deep learning model

In summary,the results revealed that the proposed model can accurately segment the weak inter layers for underconstruction rock tunnels,and exhibit a strong robustness for the small scale samples.Nevertheless,the inevitable errors and noise points caused by large size images testing are urgent to be reduced in the future study.