Deep learning-based raster digitization engine
Atul Laxman Katole, Purnaprajna Mangsuli, Omkar Anil Gune, M.S. Shakeel, Abhiman Neelakanteswara, Aria Abubakar · 2022
A vast amount of historical well log data are inaccessible to state-of-the-art artificial intelligence and machine-learning (AI/ML) techniques because they are stored as raster images in TIFF, PDF, or JPEG format. The authors propose the creation of a fully automated deep-learning-based digitization engine to transform well log data embedded in raster images into digital data that is saved in JSON, CSV, or LAS format. A typical raster image captures the well log information in its constituents as document header, tables, plot segments, depth tracks, and log header sections. The proposed hierarchical approach trains deep-learning-based raster segmentation models to first extract the plot segments, depth track, and log headers from the raster images. Another set of deep-learning-based image segmentation models is used to extract the metadata from the log header. Subsequently, a novel two-stage conditional Generative Adversarial Network (cGAN) extracts the curve pixels from the plot segments. The final digitization step transforms the extracted curves pixels to numerical logs using the extracted metadata in fully automated fashion. The curve pixel extraction achieves the normalized mean absolute error of less than 0.011 per curve in a plot segment. The manual method requires several months to digitize thousands of raster images, which now can be accomplished in a few hours with the proposed approach, achieving a more than 100 times productivity gain.