Research on the Identification Method of Slice Images of Tight Oil Reservoir Rocks Based on Improved Refine Mask

Abstract

Terrestrial tight oil exhibits strong diagenetic heterogeneity, and a large number of rock slices are required to reveal the true micro-pore throat structure characteristics. Traditional identification methods for tight oil rock slices suffer from long manual observation time, poor accuracy of machine learning method and strong subjectivity of manual judgment, making it difficult to meet the requirements of reservoir fine description and quantitative characterization. In this study, targeting the Upper Paleozoic in the North Subsag Basin of China and the Linxing Block of Ordos Basin, a deep learning-based identification method for slice features of tight oil reservoir rocks was proposed. Firstly, the image preprocessing technique was investigated and the Gaussian denoising filtering algorithm was applied to reasonably allocate Gaussian weight coefficients to the original images, ensuring the quality of the samples. Secondly, the self-labeling image data augmentation technique was constructed to address the problem of sparse samples. Thirdly, the RefineMask instance segmentation algorithm was introduced and improved to simultaneously achieve segmentation and identification of components in slices of tight oil reservoir rocks. Finally, the experiment demonstrates that the SLA-RefineMask method has significant advantages in terms of accuracy and execution speed compared to other methods.