Abstract: Aiming at the problems of low efficiency and strong subjectivity of traditional artificial joint logging methods,a residual attention full convolutional network (RC-FCN) model based on deep learning is proposed. By integrating the multiscale feature extraction ability of the fully convolutional network,the gradient optimization characteristics of the residual module,and the channel-space two-dimensional attention mechanism,a collaborative optimization architecture with cross-layer feature multiplexing and dynamic weight allocation functions is constructed. Based on the VGG16 encoder,the model introduces the ResNet residual block to enhance the expression ability of deep features,and combines the CBAM attention module to realize the accurate focusing of joint edge features,which effectively solves the problem of small-scale joint segmentation blur and background interference in complex underground scenes. The experimental results show that the RC-FCN model achieves a comprehensive recognition accuracy of 92. 5% on the downhole joint image test set,which is 7% higher than the traditional UNet model. Based on the analysis algorithm of the attitude parameters constructed by the segmentation results,the intelligent cataloging process of ′image segmentation-feature extraction-attitude calculation′ is realized. The robust characterization ability of the model to the joint geometry is verified by the sensitivity analysis of the dip angle error,which provides an efficient technical solution for the construction of intelligent mines. 

Key words: joint identification,deep learning,semantic segmentation,intelligent mine