VeerNet, a deep neural network architecture designed to address the challenges of classifying and digitizing raster well-log images.

 

 

Introduction: The paper begins by providing an introduction to the importance of well-logging in oil and gas extraction and its role in determining formation depth and oil reserves. It highlights the use of raster logs, which are scanned copies of paper logs saved as image files, as a cost-effective alternative to digital formats for preserving well-logging information. However, digitizing raster logs manually is time-consuming and inefficient. The paper proposes VeerNet as a solution to automate this process and extract valuable information from raster well-log images.

 

Previous Work: The authors review previous work in the field of well-log digitization, including commercial software and unsupervised computer vision techniques. They discuss the limitations of existing methods, which often require manual intervention and have slow processing times. The paper emphasizes the need for a more efficient and accurate approach to curve classification and digitization.

 

Proposed Approach: The authors introduce VeerNet, a deep neural network architecture designed to classify and digitize well-log curves from raster images. VeerNet follows an encoder-decoder architecture, with residual blocks in the encoder for feature extraction and transformer layers for refining the internal representation. The decoder utilizes upsampling and convolution operations to generate spatial masks indicating the presence of log curves. The proposed approach aims to balance preserving key signals while reducing dimensionality.

 

Experimental Setup: The paper describes the experimental setup used to train and evaluate VeerNet. A dataset of 10,000 images, consisting of well-log curves from LAS and raster files, is used for training. The dataset is divided into training and validation sets, and various loss functions, including Dice, Tversky, Lovasz, Focal, and Sparse Cross Entropy (SCE), are evaluated. The models are trained on multiple workers with AMD CPU cores and NVIDIA A100-SXM-80GB GPUs.

 

Results: The experimental results show that VeerNet achieves an overall F1 score of 35% and Intersection over Union of 30% in classifying and digitizing well-log curves. The best-performing loss functions are Lovasz and SCE, which yield higher F1 scores. The paper presents comparisons between the ground truth and predicted values for Gamma-ray (GR) and Caliper (CALI) logs, demonstrating the goodness of fit achieved by VeerNet. The statistical analysis shows a significant correlation between the derived values and native LAS data for the GR log.

 

Discussions: The authors discuss the limitations of the study, such as the low signal-to-noise ratio in raster logs and the need for a larger dataset for Caliper logs to improve accuracy. They propose future enhancements, including incorporating OCR architecture to automatically read scales and train the model on data from specific geologic reservoirs. The paper concludes by highlighting the potential of VeerNet as a valuable solution for digitizing raster well-log images.

 

Overall, this research paper introduces VeerNet as a deep-learning model for curve classification and digitization of raster well-log images. It presents the methodology, experimental results, and discussions, showcasing the effectiveness of VeerNet in automating the classification and digitization process. The paper contributes to the field by providing a promising solution and opening avenues for further research and improvements in well-log digitization.

 

To read the whole paper visit: Paper