关键词: 巷道围岩裂隙, 轻量化语义分割, 多尺度特征增强, 注意力机制, 智能检测

Abstract: To address the problems of computational redundancy，small target loss，and sensitivity to background interference in the traditional DeepLabv3+ model for roadway surrounding rock fracture de⁃ tection，a lightweight collaborative semantic segmentation method based on multi-scale feature enhance⁃ ment is proposed. Taking 9 mining areas of Zhaiwan Phosphate Mine as the research object，a high-resolu⁃ tion fracture dataset containing 540 samples was constructed（the fracture pixel proportion ranges from 2.7% to 4.1%）. The core of the model optimization includes four aspects: adopting MobileNetV2 as the backbone network to achieve 83% parameter compression; reconstructing the dilation rates of the atrous spatial pyramid pooling（ASPP）module（a scale feature extraction module）into［2，3，7］（coprime de⁃ sign）to enhance the receptive field of small targets; embedding the convolutional block attention module （CBAM）dual-dimensional attention mechanism after ASPP to strengthen feature localization; and introduc⁃ ing the Dice Loss-Focal Loss combined loss function（λ=0.4）to solve the problem of extreme sample imbal⁃ance. Six groups of comparative experiments were designed to verify the effectiveness of the model. The re⁃
sults show that the optimal model（M5）achieves a mean Intersection over Union（mIoU）of 74.2%，which
is 4.73 percentage points higher than that of the baseline model，and its parameter count is only 3.4 M
（about 14.8% of that of Xception）. Engineering experiment verification shows that the deviation between
the fracture parameters（such as dip angle and spacing）extracted by the model and the on-site manual
survey results is less than 8%，which can accurately support the construction of the transparent geological
model and rock mass quality classification of Zhaiwan Phosphorus Mine.This model realizes the coordinat⁃
ed optimization of fracture segmentation accuracy and lightweight performance，and can be deployed on
embedded inspection equipment. It provides reliable technical support for mine engineering geological sur⁃
veys and geological disaster early warning，and its design idea can serve as a reference for similar extreme
sample segmentation tasks.

Key words: roadway surrounding rock fractures, lightweight semantic segmentation, multi-scale fea? ture enhancement, attention mechanism, intelligent detection