Research On Intelligent Identification And Scale Characterization Algorithm Of Pore And Fissure Of Loess Based On Image Segmentation

Since the implementation of the " One Belt and One Road" national strategy,major projects such as ditch management,land transformation,whittle the mountains and build the land have been rapidly advancing in the Loess Plateau region of the western Silk Road.While benefiting the people,they have also caused disturbances to the fragile Loess Plateau environment and frequent landslide disasters.It is the basis to explore the disaster driving mechanism and evolution law by in-situ detection of loess pore cracks to obtain the macro-fine-micro-structure information and spatial distribution of loess.However,there is still a lack of reliable technical means.The pipeline robot equipped with vision system provides a possibility to solve this problem.Based on this,this paper aims to systematically study the image recognition algorithm and scale characterization method in in-situ detection of loess pore cracks.The main work is as follows:Firstly,the construction of a dataset that fully reflects the structural characteristics of loess pore-cracks is a necessary condition for developing intelligent image recognition algorithms,but there is a lack of relevant datasets both domestically and abroad.Therefore,a dedicated data collection platform was developed.Based on the analysis of the environmental characteristics of loess pore structure in geological exploration,the spatial distribution characteristics of loess structure,factors affecting the quality of dataset production,and the limitations of training data volume in deep learning,undisturbed loess at different depths in different regions were selected,and image acquisition planes were prepared to collect 1232 images,which were manually calibrated to construct the dataset.Secondly,aiming at the extraction of structural morphological information,an intelligent recognition algorithm LOE-Net(Loess Identification Network)for loess pore-crack was constructed based on image segmentation technology to extract pore-crack structures in loess images.Swin-unet network architecture was selected for its structural characteristics,and Lawin ASPP module was integrated in the decoding part to fuse multi-scale features.Dice loss and Focal loss weighted combination was used as the loss function,and Lovasz-Softmax loss was used for refinement to solve the problem of foreground-background imbalance and category imbalance in the dataset.Compared with other authoritative algorithms,the feasibility and advancement of the LOE-Net algorithm constructed in this paper in the intelligent recognition of loess pore-cracks were proven.Thirdly,a scale characterization algorithm for intelligent recognition results of loess pore-cracks was constructed to meet the demand for pore-crack structure parameters in loess engineering.Connected region analysis was used to extract pore,micropore,and crack structures in loess images,and the structures were restored to world coordinate system scale through camera calibration experiments.Quantitative parameters were used to represent the structural size and morphology of loess pore-cracks,which meet the practical needs of engineering.Finally,the engineering feasibility of LOE-Net and the accuracy of the scale characterization algorithm were verified through experiments.Field experiments were conducted in the comprehensive model experimental base of Weinan City,Shaanxi Province,using a pipeline robot with a visual system to photograph the wall images of loess holes,and LOE-Net was used to recognize the pore and crack structure of the wall images,which verified the engineering feasibility of LOE-Net.In the laboratory,loess cutting plane images were recollected,and LOE-Net was used for recognition and the scale characterization algorithm was used to calculate pore and crack sizes,which verified the accuracy of the intelligent recognition and characterization algorithm for loess pore and crack structures.