Rock Pores Segmentation And Reconstruction Based On Deep Learning

Rock pores are an extremely important part of the rock composition and the main controlling factor for the study of oil and gas reserves.The accurate segmentation of rock pores and the reconstruction of the 3D pore model are not only helpful for deeper research on pores,but also play a vital role in predicting the productivity of oil and gas reservoirs and improving oil and gas recovery.The existing traditional pore segmentation method has been widely used,but it is still susceptible to the interference of rock image noise to affect the segmentation,and it is difficult to segment the small and narrow pores in the rock.There are relatively few rock data sets currently available,which affects the study of rock pores to a certain extent.In response to the above problems,the main research work carried out in this paper is as follows:First of all,in view of the problem of fewer existing rock data sets,a set of a tight sandstone core data set in Jilin including horizontal slices and vertical slices was produced,and the data was processed by a combination of histogram equalization and median filtering processing,increase the contrast between the aperture and the background pixel,reduce the interference of image noise,enrich the aperture information,and lay the data foundation for the subsequent image segmentation.Then,in view of the small and long pores prone to leaking segmentation during the segmentation process,the channel attention module is introduced to increase the weight of such pores and reduce the leaking segmentation.Constructing recurrent residual convolution module based on the dilated convolution to help the network fits more rock pore features and improves the segmentation accuracy.The UNet deep learning network and the above modules are combined as the final pore segmentation network model AR-UNet(Attention and Recurrent Residual Convolution UNet)in this paper,and the pooling index is introduced into the sampling stage of the network.Finally,the pore results segmented by the model in this paper are compared with other deep learning network models,and the model in this paper has better segmentation results.Setting up an ablation experiment proved the effectiveness of the attention mechanism module and the cyclic residual convolution module,which can solve the problem of difficult segmentation of small and long pores in the rock.Finally,In order to reconstruct the pore 3D model more accurately,the horizontal pore results of the network segmentation and the vertical pore results are fused to synthesize the final segmented pore results,then use the hierarchical quadtree to store the data and based on the fusion results,the pore 3D model is reconstructed by the voxel method.During the reconstruction,using the single element model plus matrix drawing method to draw,to achieve the dual optimization of memory and video memory.The pore three-dimensional model reconstructed by the segmentation result can also reflect the pore structure more rationally and intuitively.