Lithology Identification using Well Logging Images Based on Improved Inception Network

The lithology identification plays a fundamental role for the fine characterization and comprehensive evaluation of complex reservoirs. Several methods so far have been proposed for lithology identification such as the analysis of sensitive conventional logging curves, intersection diagrams as well as geostatistical analysis. Though these methods can lead to improving accuracy of lithology recognition, they are either indirectly reflect lithology information or affected by handwork factors, thus the efficiency and accuracy are limited. Recently, in the image processing area, Inception convolutional neural network has become a powerful tool for expressing complex structures and extracting multi-scale feature information due to its parallel structure of convolutional kernel. To this end, in this paper, we propose to apply the Inception network for lithology identification based on high-resolution well-logging imaging. Different from the conventional Inception network, the proposed model is able to automatically extract the typical feature regions of different lithology categories from the raw Formation MicroScanner Image (FMI) data by parameter transfer strategy, which could improve the generalization performance and enhance the accuracy of lithology identification. To alleviate the problem of differences in the distribution of features, the improved Inception model adopts regularized loss function constrained full connection layer to extract high-level features and obtain more available and valuable information. Experimental results show that the network model has the advantages of fewer parameters and less computational load, and the accuracy of lithology identification can reach 97.32%.