Research On Video Stitching Method Of Geotechnical Drilling Based On Clear Perception Mechanism

Drilling is a routine measure to understand the state of the surrounding rock during underground construction.By collecting and analyzing the video of the borehole wall in the borehole,we can fully understand the fracture development of the surrounding rock,and then evaluate the safety of the project.For the traditional borehole video,researchers observe the video frame by frame and mark the cracked or fractured areas manually.However,due to the large amount of data and low resolution of borehole images,this manual analysis method is inefficient and error-prone,and it is difficult to obtain objective and accurate analysis results.Therefore,this paper proposes a geotechnical borehole video processing method based on clear perception mechanism,which expands the collected borehole wall images sequentially and stitches them together effectively,thus enabling researchers to analyze the borehole fracture situation intuitively.The specific studies are as follows.(1)A new Gamma Transform Enhanced Hough Gradient Circle Recognition Model(GHGC)is proposed to address the problem of inaccurate recognition of borehole circles due to the influence of illumination.The GHGC model is developed by integrating the The GHGC model improves the recognition of circle centers under different illumination intensities by incorporating the Gamma transform into the Hough gradient circle recognition model.It is proved that the Hough gradient circle algorithm improves the recognition ability of circle centers for images with different exposure degrees by fusing the Gamma transform,which also reduces the number of parameters,and improves the computational efficiency.(2)An Autoencoder-based Feature Extraction Model(A-FEM)is proposed for image distortion due to blurring,weak texture and unfolding of the image itself.The approach of this paper is to incorporate the autoencoder module into a coarse-to-fine feature extraction network.The network uses the autoencoder to compute the appropriate convolution kernel for each scale of the image,thus reducing the input length of the next stitching model and improving the accuracy of feature point matching.It is demonstrated that,compared with other feature extraction networks,the A-FEM network is more perceptive for low-resolution borehole images due to the incorporation of auto-encoder,which results in high semantic and accurate feature maps at different scales,reflecting the advantages of the A-FEM network on lowresolution borehole images.(3)A Boosting Vision Transformer Model for Low-resolution Borehole Image Stitching through Algebraic Multigrid(AMG&VITM)is designed to address the problem of stitching seams and black pixel blocks caused by low-resolution borehole image stitching.Image Stitching through Algebraic Multigrid,AMG&VITM).The method firstly segments the image,records the local pixel values and encodes the position,then calculates the MSE value of each image block using the algebraic multigrid,and passes the filtered image blocks into the Transformer encoder by thresholding,and finally gets the coordinates of the two images matching by MLP,and stitches the image using the affine transform.It is proved that the model identifies the matched feature points and calculates the vertical displacement difference to determine the final stitching area by combining the affine transform,which makes it overcome the problems of stitching seams and black distorted areas caused by image distortion and blurring.To test the performance of the proposed model and network in this paper,we collected videos from different geotechnical projects and split them into 170,000 low-resolution borehole images as our dataset,such as ore,oil and gas projects.The experimental results show that the model and network can better solve the low-resolution borehole image stitching problem and obtain good visual effects,which can make contribution to the safety analysis of geotechnical projects.