Research On Dynamic Identification Method Of Rock Cracks Based On Convolutional Neural Networks

Nowadays,with the rise of deep learning,image processing using computer vision to solve related problems has become a hot spot for experts in various fields at home and abroad.The combination of deep learning and civil engineering has gradually become one of the hot research directions in the future.In this paper,by establishing multi-source rock fracture data set and combining with the image features of rock fracture,the classical convolutional neural network is improved to realize the dynamic identification of rock fracture,and the automatic calculation of the fracture area is carried out according to the identification results.The main research contents are as follows:(1)The relevant theories and research status of deep learning and convolutional neural networks are briefly reviewed.The components of convolutional neural network are summarized in detail,and their functions and significance are introduced respectively.Several common deep learning frameworks are introduced and compared,and Tensorflow is selected as the deep learning framework for this study according to their characteristics and the research requirements of this paper.Provide theoretical support for the follow-up research.(2)The U-net structure is improved based on the features of rock fissure image,and the Trans UNet algorithm is obtained by combining U-net and Vision Transformer algorithm,and its applicability and shortcomings for rock fissure segmentation are analyzed.By adding the attention mechanism,optimizing the coding structure of VIT algorithm and optimizing the convolutional layer of decoder,Thus,an improved Trans UNet algorithm suitable for rock fracture image segmentation is obtained.In view of the lack of open data set for existing rock fractures,the CT images of rock fractures collected by the network,the crack images of granite test blocks prepared by the laboratory and the crack images of natural mountain rock fractures based on aerial photography of unmanned aerial vehicle were collected and expanded by data enhancement method.labelme was used for fine annotation.Thus,the multi-source rock fracture data set required for subsequent model training is obtained.(3)Set up the experimental environment,install the required library and framework,input the established data set and adjust the training parameters,and obtain the semantic segmentation model based on the improved Trans UNet algorithm and rock fracture recognition.The performance of the model is compared with the performance of five classical and representative semantic segmentation algorithms by using relevant evaluation indexes,proving the superiority of the model.At the same time,two other kinds of crack images outside the range of data set were selected to verify the generalization ability of the model.Through uniaxial compression test of granite standard test block and shooting video,the video file of its failure process is obtained.Then,Open CV library is used to write programs to recognize the video of its failure process,so as to realize dynamic recognition.Moreover,according to the principle of pixel statistics,the area of fracture region in rock fracture image and dynamic video is automatically calculated.The current crack area and crack rate are displayed in real time.