相似材料模型实验是研究矿山开采沉陷及损害的重要手段，实验中高精度地获取模型变形信息至关重要。现有的多种模型变形测量方法在数据采集精度和效率方面仍存在一定的局限性。本文以陕西彬长文家坡煤矿4^#煤层4105工作面为模型开展相似材料模拟实验，利用非量测相机近距离拍摄模型得到高分辨率的序列图像，通过图像畸变纠正、自动拼接及改进的Faster R-CNN(Faster Regions with Convolutional Neural Network Features)算法进行特征点提取，高精度地获取模型开采过程中的变形信息。主要研究内容及结果如下：

（1）通过对比分析相似材料模型变形测量中不同近景摄影测量方法的特点，采用一种类航摄模式的序列图像采集技术方案。结果表明，现有方法中全景固定成像耗时短，但精度较低；单点数码成像精度较高，但外业拍摄和数据处理用时长；数字近景工业摄影测量系统可高效作业，但对于设备要求高，不具有普适性。采用本文的类航摄序列图像采集方法在作业效率、图像分辨率和测量精度方面均具有明显优势。

（2）针对相似材料实验模型特点，在分析测点标志、大小、像控点数量、拍摄物距、图像重叠度对模型测量精度影响的基础上，确定了实验模型类航摄一体化作业的技术流程。模型开挖过程中采用非量测相机进行近距离正射拍摄，形成横向重叠度大于70%，竖向重叠度大于50%的图像序列，利用双线性插值方法进行图像畸变校正处理，运用全自动稳健的图像拼接算法，使用Harris角点检测算子进行特征点提取，利用RANSAC算法对伪匹配点进行分类处理，通过颜色插值法进行接边处理，进一步对图像序列进行拼接。结果表明，所生成的全景图像分辨率相比固定相机姿态一次拍摄模型全景图像提高约10倍。

（3）在序列图像处理中提出改进的Faster R-CNN算法，分析了该算法对于改进模型测点提取精度的实际效果。在图像处理过程中以测点标志为训练网络，基于Faster R-CNN算法加入置信度，通过选取合适的滑动模板进行区域分析，再利用相关性的图像匹配算法得到测点精确坐标。结果表明，该方法使得检测正确率提升4.7%，检测效率提升2.3%，而模型实验中测点坐标提取的精度优于0.027mm，相当于实地测量精度5.4mm，满足矿山相似材料模型实验要求。

（4）以文家坡煤矿4105工作面地质采矿条件制作相似材料模型，利用本文方法提取的变形信息，初步揭示了该工作面开采引起的岩层与地表变形特征。制作比例尺1:200的模型进行17次模拟开挖，获取17期图像序列并利用改进的Faster R-CNN算法进行特征点提取和变形分析。结果表明，模型地表下沉曲线与实地观测数据基本相符，最大下沉量达-3.767mm，相当于实地-753mm。通过模型实验也揭示了覆岩中周期来压和冒落、裂隙带发育的基本特征。研究结果对于开采区地面水库的安全防护具有一定的参考价值。

Similar material model experiment is an important means to study mining subsidence and damage, and it is very important to obtain model deformation information with high precision in the experiment. There are still some limitations in the accuracy and efficiency of data acquisition of existing model deformation measurement methods. In this paper, the 4105 working face of No.4 coal seam in Wenjiapo Coal Mine, Binchang, Shaanxi Province is used as the model to carry out the similar material simulation experiment. The high-resolution sequence images are obtained by taking a close shot of the model with a non-metric camera, and the feature points are extracted by image distortion correction, automatic stitching and improved Faster R-CNN (Faster Regions with Convective Neural Network Features) algorithm, so that the deformation information in the mining process of the model can be obtained with high precision. The main research contents and results are as follows:

(1) By comparing and analyzing the different characteristics of four close-range photogrammetry methods existing in deformation measurement of similar material models, a technical scheme of sequence image acquisition with similar aerial photography mode is adopted. The results show that the existing panoramic fixed imaging methods take short time, but the accuracy is low. The precision of single-point digital imaging is high, but it takes a long time for field shooting; Digital close-range industrial photogrammetry system can work efficiently, but it has high requirements for equipment and can't display the deformation of specific points. The aerial photography-like sequence image acquisition method adopted in this paper has obvious advantages in operation efficiency, image resolution and measurement accuracy.

(2) According to the characteristics of experimental models of similar materials, based on the analysis of the influence of measuring point mark, size, number of image control points, object distance and image overlap on the measurement accuracy of the model, the technical process of aerial photography integrated operation of experimental models is determined. In the process of model excavation, non-metric camera was used to take close-range orthophoto shots, forming image sequences with horizontal overlap greater than 70% and vertical overlap greater than 50%. Bilinear interpolation method was used to correct image distortion, automatic and robust image stitching algorithm was used, Harris corner detection operator was used to extract feature points, RANSAC algorithm was used to classify pseudo-matching points, and color interpolation method was used to join edges, so as to further splice the image sequences. The results show that the resolution of the generated panoramic image is about 10 times higher than that of the panoramic image of the fixed camera.

(3) An improved Faster R-CNN algorithm is proposed in sequence image processing, and the actual effect of this algorithm on improving the accuracy of measuring point extraction of the model is analyzed. In the process of image processing, the measuring point mark is used as the training network, the confidence level is added based on Faster R-CNN algorithm, the appropriate sliding template is selected for region analysis, and then the accurate coordinates of the measuring point are obtained by using the correlation image matching algorithm. The results show that this method can improve the detection accuracy by 4.7% and the detection efficiency by 2.3%, and the accuracy of coordinate extraction of measuring points in model experiment is better than 0.027mm, which is equivalent to the field measurement accuracy of 5.4mm, which meets the requirements of model experiment of similar materials in mines.

(4) Based on the geological mining conditions of No.4105 working face in Wenjiapo Coal Mine, a similar material model is made, and the deformation information extracted by this method is used to preliminarily reveal the characteristics of strata and surface deformation caused by mining in this working face. A 1:200 scale model was made for 17 simulated excavations, and 17 image sequences were obtained. The improved Faster R-CNN algorithm was used for feature point extraction and deformation analysis. The results show that the surface subsidence curve of the model is basically consistent with the field observation data, and the maximum subsidence is -3.767mm, which is equivalent to -753mm in the field. The model experiment also reveals the basic characteristics of periodic weighting and caving and fracture zone development in overburden. The research results have certain reference value for the safety protection of surface reservoirs in mining areas.