基于机器学习的石窟顶板裂隙热红外识别

Machine-learning-based thermal infrared recognition of fractures in grotto roofs

石窟顶板层状岩体中发育的裂隙相互交切,极易引发石窟岩体的失稳破坏,对其快速精准识别是石窟保护的重要基础。针对石窟顶板岩体裂隙的非接触精准测量需求,结合热红外探测技术和改进的UNet网络模型,对顶板裂隙网络二值图进行提取,并运用聚类算法,完成了裂隙网络二值图分割识别以及裂隙分组。结果表明,该网络模型各项性能相较于其他网络模型有所提高,Dice系数和推理速度分别达到了71.63%和0.84帧/s,识别过程抵抗人工结构物影响的能力较强,凸显该方法推理速度快,提取精度高、热红外图像适用性好等特点。以安岳圆觉洞顶板为例,应用该方法共分割识别出153条裂隙,并确定了NW327°和NE55°是顶板裂隙的优势走向,与其他测量方法相比识别效果更好。

The cracks developed in the layered rock mass of grotto roofs intersect with each other,which can easily cause instability and failure of the cave rock mass.Rapid and precise fracture identification is crucial for grotto protection.To meet the need for non-contact,precise fracture measurement,this study integrates thermal infrared detection technology with an improved UNet network model to extract binary maps of roof fracture networks.Clustering algorithms are employed for segmentation and recognition,achieving a Dice coefficients of 71.63%and a detection speed of 0.84 frames/s.The method exhibits high extraction efficiency,accuracy,good applicability of thermal infrared images and resilience against artificial structure influence.Applied to the roof of Anyue Yuanjue Grotto,this method successfully identified 153 fractures and reveals dominant fracture trends at NW327°and NE55°,outperforming other measurement techniques.