Digitization of Raster Well Logs using Multimodal Prompting and Computer Vision Techniques

This work presents an automated methodology for digitizing raster well logs using multimodal prompting and advanced computer vision techniques. Raster well logs, commonly stored as high-resolution TIFF images, contain critical geological data plotted as curves like Gamma-ray, resistivity, and caliper curves. Manual digitization of these logs is labor-intensive and costly, especially for large datasets. Our approach leverages cutting-edge models like Visual Question Answering (VQA) and the Segment Anything Model (SAM) to extract user-specific curves efficiently.

Key steps include preprocessing to separate header and plot regions, extracting curve properties via VQA, and performing text-guided segmentation with SAM to generate high-quality binary masks. Gridline removal and depth scale normalization enhance accuracy, enabling the digitized curves to be exported in LAS format. Tested on a dataset of raster logs, our method achieved an average normalized root mean squared error (NRMSE) of 0.167, demonstrating near-perfect reconstruction.

This solution drastically reduces the time and cost of well log digitization, offering a scalable alternative to manual methods. Future work will address challenges like handling multiple curves with varying scales within the same image.