Journal of Petroleum Research

Journal of Petroleum Research

A Comparative Study of Convolutional Autoencoder and Vision Transformer Architectures for Denoising and Reconstruction of Subsurface Fractured Reservoir Rock Images

Document Type : Research Paper

Author
Behrad Tabrizipour
Faculty of Petroleum Engineering, Science and Research Branch, Islamic Azad University, Tehran, Iran
10.22078/pr.2026.5787.3573
Abstract
Fractured reservoir rocks are among the most critical targets in petroleum engineering studies, as their complex networks of fractures and cracks play a decisive role in the porosity and permeability of these formations. CT-scan imaging is a key tool for analyzing the internal structure of such rocks; however, reduced image quality often makes accurate reconstruction challenging. In this study, two deep-learning architectures a Convolutional Autoencoder and a Vision Transformer were designed and compared with the objective of enhancing and reconstructing CT-scan images of fractured reservoir rocks. The primary task of the models was to reconstruct and denoise subsurface fractured images, and following reconstruction, a three-class segmentation process (matrix, open fracture, and filled fracture) was performed using the classical Otsu thresholding method to evaluate the improved performance of the reconstructed outputs. Quantitative and qualitative results demonstrated that the Vision Transformer outperformed the autoencoder due to its use of the attention mechanism. The Vision Transformer achieved a peak signal-to-noise ratio of 45 and a structural similarity index of 0.98 in image reconstruction, whereas the autoencoder achieved values of 39 and 0.93, respectively. In the Otsu-based three-class segmentation, the Vision Transformer again delivered superior performance in terms of precision, recall, and F1-score. For the matrix class, precision, recall, and F1-score reached 0.989, 0.990, and 0.989, compared to 0.980, 0.969, and 0.974 for the autoencoder. For the filled-fracture class, the results were 0.974, 0.953, and 0.963, and for the open-fracture class, 0.923, 0.920, and 0.921, while the autoencoder recorded 0.960, 0.962, 0.961 and 0.940, 0.934, 0.937 for the same classes. Overall, the findings indicate that the Vision Transformer is a more efficient and accurate option for CT-scan image reconstruction of fractured rocks, offering significant potential to improve numerical analyses of rock properties such as porosity and permeability in reservoir engineering studies.
Keywords
Fractured reservoir rock
image reconstruction
autoencoder
vision transformer
peak signal-to-noise ratio
structural similarity index
Subjects
[1]. Nguyen, K. L., Jumah, M., Verma, N., & Quttainah, R. (2023). Identifying Fracture Sweet Spots Using Artificial Neural Network Approach: A Case Study in Najmah/Sargelu Reservoir, Kra Al-Maru Field, West Kuwait. Middle East Oil, Gas and Geosciences Show, doi.org/10.2118/213770-MS.
[2]. Wilson, T. H., Smith, V., & Brown, A. (2015). Developing a model discrete fracture network, drilling, and enhanced oil recovery strategy in an unconventional naturally fractured reservoir using integrated field, image log, and three-dimensional seismic data. AAPG Bulletin, 99(4), 783–808. doi.org/10.1306/10031414015.
[3]. Sun, S., & Pollitt, D. A. (2021). Optimising development and production of naturally fractured reservoirs using a large empirical dataset. Petroleum Geoscience, 27(3), petgeo2020-079. doi.org/10.1144/petgeo2020-079.
[4]. Lyu, W., Zeng, L., Zhou, S., Du, X., Xia, D., Liu, G., Li, J., & Weng, J. (2019). Natural fractures in tight-oil sandstones: A case study of the Upper Triassic Yanchang Formation in the southwestern Ordos Basin, China. AAPG Bulletin, 103(1), 135–154. https://doi.org/10.1306/0130191608617115.
[5]. Hashemi, R., Saberi, F., Asoude, P., & Soleimani, B. (2024). Enhancing Reservoir Zonation through Triple Porosity System: A Case Study. SPE Journal, 29(06), 3043–3062. doi.org/10.2118/219491-pa. [6]. Al-Rbeawi, S. (2020). The performance of complex-structure fractured reservoirs considering natural and induced matrix block size, shape, and distribution. Journal of Natural Gas Science and Engineering, 81, 103400. doi.org/10.1016/j.jngse.2020.103400.
[7]. Zhu, Y., Hou, D., Lan, M., Sun, X., & Ma, X. (2019). A comparison of ultra-high-resolution CT target scan versus conventional CT target reconstruction in the evaluation of ground-glass-nodule-like lung adenocarcinoma. Quantitative Imaging in Medicine and Surgery, 9(6), 1087–1094. doi.org/10.21037/qims.2019.06.09.
[8]. Faisal, R. S., Salih, N. M., Kamal, I., & Préat, A. (2023). X-ray computed tomography (CT) to scan the structure and characterize the mud cake incorporated with various magnetic NPs concentration: An application to evaluate the wellbore stability and formation damage. Nanomaterials, 13(12), 1843. doi.org/10.3390/nano13121843.
[9]. Cappuccio, F., Toy, V. G., Mills, S., & Adam, L. (2020). Three-dimensional separation and characterization of fractures in X-ray computed tomographic images of rocks. Frontiers in Earth Science, 8, 529263. doi.org/10.3389/feart.2020.529263.
[10]. Soltanmohammadi, R., & Faroughi, S. A. (2023). A comparative analysis of super-resolution techniques for enhancing micro-CT images of carbonate rocks. Applied Computing and Geosciences, 20, 100143. doi.org/10.1016/j.acags.2023.100143.
[11]. Bizhani, M., Haeri Ardakani, O., & Little, E. (2022). Reconstructing high fidelity digital rock images using deep convolutional neural networks. Scientific Reports, 12, Article 4264. doi.org/10.1038/s41598-022-08170-8.
[12]. Shahmoradi, M., Rekabdar, G., & Adibi, S. (2016). A comparative study of new and current methods for dental micro-CT image denoising. Dentomaxillofacial Radiology, 45(3), 20150302. doi.org/10.1259/dmfr.20150302.
[13]. Feng, X., Zhang, W., Su, X., & Xu, Z. (2021). Optical Remote Sensing Image Denoising and Super-Resolution Reconstructing Using Optimized Generative Network in Wavelet Transform Domain. Remote. Sens., 13, 1858. doi.org/10.3390/rs13091858.
[14]. Sahu, P., Kumar, A., & Rout, S. (2022). Wavelet guided 3D deep model to improve dental microfracture detection. In Applied Medical Artificial Intelligence (pp. 1–12). Springer. doi.org/10.1007/978-3-031-17721-7_16.
[15]. Mathew, E. S., Singh, K., & Arbogast, B. (2025). Deep learning assisted denoising of fast polychromatic X-ray micro-CT imaging of multiphase flow in porous media. Computers & Geosciences, 204, 105990. doi.org/10.1016/j.cageo.2025.105990.
[16]. Koetzier, L. R., Mastrodicasa, D., Szczykutowicz, T. P., van der Werf, N. R., Wang, A. S., Sandfort, V., van der Molen, A. J., Fleischmann, D., & Willemink, M. J. (2023). Deep Learning Image Reconstruction for CT: Technical Principles and Clinical Prospects. Radiology, 306(3), e221257. doi.org/10.1148/radiol.221257.
[17]. Janssens, N., Huysmans, M., & Swennen, R. (2020). Computed Tomography 3D Super-Resolution with Generative Adversarial Neural Networks: Implications on Unsaturated and Two-Phase Fluid Flow. Materials (Basel, Switzerland), 13(6), 1397.doi.org/10.3390/ma13061397.
[18]. Shirmard, H., Farahbakhsh, E., Heidari, E., Beiranvand Pour, A., Pradhan, B., Müller, D., & Chandra, R. (2022). A comparative study of convolutional neural networks and conventional machine learning models for lithological mapping using remote sensing data. Remote Sensing, doi.org/10.3390/rs14040819.
[19]. Liu, L., Chang, B., Prodanović, M., & Pyrcz, M. J. (2025). AI-based digital rocks augmentation and assessment metrics. Water Resources Research, 61, e2024WR037939. doi.org/10.1029/2024WR037939.
[20]. Espinoza, D. N., Shovkun, I., Makni, O., & Lenoir, N. (2016). Natural and induced fractures in coal cores imaged through X-ray computed microtomography: Impact on desorption time. International Journal of Coal Geology, 154-155, 165–175. doi.org/10.1016/j.coal.2015.12.012.
[21]. Hinton, G. E., & Salakhutdinov, R. R. (2006). Reducing the dimensionality of data with neural networks. Science, 313(5786), 504–507. doi.org/10.1126/science.1127647.
[22]. Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). Learning representations by back-propagating errors. Nature, 323(6088), 533–536. doi.org/10.1038/323533a0.
[23]. Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., & Houlsby, N. (2020). An image is worth 16×16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929. arxiv.org/abs/2010.11929.
[24]. Huynh-Thu, Q., & Ghanbari, M. (2008). Scope of validity of PSNR in image/video quality assessment. Electronics Letters, 44(13), 800–801. doi.org/10.1049/el:20080522.
[25]. Wang, Z., Bovik, A. C., Sheikh, H. R., & Simoncelli, E. P. (2004). Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4), 600–612. doi.org/10.1109/TIP.2003.819861.
[26]. Tatbul, N., Lee, T. J., Zdonik, S., Alam, M., & Gottschlich, J. (2018). Precision and recall for time series. In Proceedings of the 32nd Conference on Neural Information Processing Systems (NeurIPS 2018), Montréal, Canada. doi.org/10.48550/arXiv.1803.03639.
[27]. Foody, G. M. (2023). Challenges in the real world use of classification accuracy metrics: From recall and precision to the Matthews correlation coefficient. PLOS ONE, 18(10), e0291908. doi.org/10.1371/journal.pone.0291908.
[28]. Wang, Z., Ning, X., & Blaschko, M. B. (2023). Jaccard metric losses: Optimizing the Jaccard index with soft labels. Proceedings of the 37th Conference on Neural Information Processing Systems (NeurIPS 2023). doi.org/10.48550/arXiv.2302.05666.
[29]. Sobhani Oghaz, A., & Emami Niri, M. (2022). Effect of fracture geometrical properties on flow behavior and dynamic sensitivity analysis of fracture parameters in fractured reservoirs. Petroleum Research, 127, 49–64. doi.org/10.22078/PR.2022.4893.3188.
Journal of Petroleum Research
Volume 36, Issue 1 - Serial Number 146
May and June 2026
Pages 102-126