Intelligent characterization of particles and pores in thin slice images of tight reservoirs based on deep learning algorithm

In the thin-section image analysis of tight sandstone reservoir, to solve the problems such as low accuracy and heavy work of traditional methods, TransUnet and Unet neural networks by combining Transformer with convolutional neural network(CNN) are used for efficient characterization of particles and pores. The TransUnet has excellent performance in particle characterization. The experiment shows that the intersection over union(IoU) reaches 0.86, with the recall rate of 0.824 and precision of 0.839, which is superior to traditional methods, proving its effectiveness in tight particle segmentation. The Unet shows efficient characterization of pores as well, with the IoU of 0.824, recall rate of 0.843 and precision of 0.953. Besides, experiment indicates that although porosity affects IoU, the model still maintains high efficiency and accuracy generally. These results fully demonstrate that deep learning method, especially TransUnet, is significantly effective in accurate segmentation of thin section images of complex tight reservoir, providing new ideas for the study of unconventional tight reservoir and showing its great potential in the field of geology.