A Study On Image Segmentation Method For Thin Section Of Tight Sandstone Reservoir Rock Based On U-Net

Image segmentation of tight sandstone is an important task in geology.This task can help researchers accurately characterize the pore structure of reservoirs and fluid migration mechanisms,greatly reducing exploration and development difficulties.It has important application value for the evaluation and exploration of unconventional oil and gas reservoirs.Traditional methods rely on manual visual inspection to segment images of tight sandstone,which suffer from problems such as subjectivity and low efficiency.Therefore,the research on automatic segmentation methods for images of tight sandstone can improve the efficiency of image segmentation,save labor costs,and have important significance for the development of this field.In recent years,deep learning algorithms have made significant achievements in the field of image segmentation.However,there are two main limitations to such algorithms: firstly,deep learning-based image segmentation models require a large number of training samples to learn as many discriminative features as possible.However,due to the difficulty of data acquisition,high annotation costs,and privacy protection issues,tight sandstone image samples are scarce and difficult to meet the training requirements of deep learning algorithms.In addition,traditional convolutional neural networks have shortcomings in extracting global image information and long-range dependency relationships,which can easily lead to inaccurate component recognition or low segmentation accuracy.To address these issues,this paper proposes a tight sandstone image augmentation segmentation model based on semantic segmentation methods,combined with data deformation methods and data oversampling image augmentation methods.The paper provides a detailed analysis and implementation of the model composition.The main research contents of this paper include the following three aspects:1.To address the problems of low data volume,insufficient diversity and high annotation cost of tight sandstone images,this paper proposes a hybrid data augmentation method combining Auto Augment and Style Generation Adversarial Network(Style GAN).Firstly,to improve the data diversity,we improve the style control method and the adaptive augmentation intensity adjustment period of the style generation adversarial network,and propose the selfattentive style generation adversarial network(SA-Style GAN)to generate high-quality tight sandstone images.Secondly,in order to expand the data scale,the search space of the augmentation strategy of the automatic augmentation algorithm is redesigned and the search algorithm is improved,and the automatic data augmentation algorithm based on the adaptive stochastic natural gradient method(ASNG-AA)is proposed to automatically search for the optimal augmentation strategy and realize the data augmentation with labels.Finally,to verify the effect of the augmentation,U-Net is used as the validation algorithm,and the method is proved to have obvious advantages in generating image quality and improving model generalization ability by designing comparison tests.2.A U-Net tight sandstone image segmentation network(CBAM-Transformer U-Net,CTU-Net)based on CBAM-Transformer is proposed for the problem of insufficient ability to extract global information and long-distance dependencies from traditional convolutional neural network images.First,the channel and spatial attention module(CBAM)is introduced into the U-Net network to enhance the model’s ability to perceive global contextual information.Second,the Transformer module is added to the decoder to help the model better capture longdistance dependencies.In addition,the Dice loss function is used instead of the traditional crossentropy loss function to reduce segmentation errors caused by data imbalance,thus improving the segmentation accuracy of the model.Finally,to verify the performance of the model,CTUNet and other comparison algorithms are trained using the augmented tight sandstone images,and segmentation experiments are performed on the test set.The experimental results show that the CTU-Net model exhibits the best segmentation performance.3.The tight sandstone image segmentation system is designed and implemented for the demand of tight sandstone image segmentation business in geological field.The system is based on CTU-Net image segmentation model,and the effectiveness of the proposed method is verified through practical application.Compared with three U-Net based semantic segmentation methods,experimental results show that the proposed tight sandstone image augmentation segmentation model in this paper achieves significant improvements in recognition and segmentation performance.The recognition accuracy of each component is above 87%,and the segmentation error rate is below16%,which can effectively realize the recognition and segmentation of tight sandstone images.The tight sandstone image segmentation system can quickly and automatically segment images,analyze the segmentation results,and greatly improve work efficiency.Additionally,the model can effectively reduce subsequent problems such as geological feature misjudgments and reservoir prediction errors caused by incorrect segmentation,and has high practical value.