Intelligent Pore Recognition Method for Carbonate Rock Electrical Image Logs Based on Deep Learning

Electrical image logs can intuitively reflect the development status and characteristics of dissolution pores, which is of significant importance for the development of oil and gas resources. However, traditional methods for identifying pores in electrical image logs are not only cumbersome and labor-intensive but also incapable of distinguishing between different types of pores. Moreover, the strong heterogeneity and dissolution effects in carbonate reservoirs result in significant variations in pore size and complex, diverse pore morphologies, making it difficult to extract pore parameters. To address these issues and challenges, this paper proposes a semantic segmentation model, FILnet, designed using computer vision technology and deep learning frameworks. This model aims to achieve intelligent recognition and segmentation annotation of pores of different scales in the wellbore region of electrical image logs. The data selection process involved using a sliding window to choose electrical log images containing dissolution pores and caves. Image processing techniques were then applied to complete and augment the images, thereby enhancing data diversity. Furthermore, a dual-attribute dataset was created using dynamic and static images from electrical image logs to assist the model in learning the semantic features of pores. Finally, the proposed model was compared with traditional pore identification methods, such as threshold segmentation. The results showed that FILnet demonstrated significant performance advantages on the dual dataset, with a mean intersection over union (MIoU) of 85.42% and a pixel accuracy (PA) of 90.54%. Compared to traditional pore identification methods, the deep learning semantic segmentation approach not only achieves recognition of different types of pores but also improves identification accuracy. This indicates that the network model and data processing methods proposed in this paper are effective and can achieve intelligent recognition and accurate segmentation of pores in electrical image logs.