Modification of the U-Net network configuration for rock sample classification and segmentation tasks

Currently, there is a steady trend towards the introduction of artificial intelligence in all technological areas. It has not bypassed the oil and gas industry where machine learning and, in particular, neural networks are now actively used. An analysis of domestic and foreign works in this area was conducted. In particular, methods and algorithms based on artificial intelligence that work with rock samples were considered. The result of this analysis was the conclusion that the use of neural networks to solve the problem of classification and segmentation of samples is becoming increasingly widespread. In addition, the reviewed works identified the limitations of the used architectures that negatively affect the efficiency of the resulting models. These limitations became the problem solved in this study. This work is based on the problem of modifying the U-Net network architecture to improve the efficiency of work with rock samples. As a modification, it is proposed to: add intermediate connections between the decoder layers for a more accurate transfer of features of varying degrees of abstraction; introduce residual blocks to solve the problem of gradient attenuation; replace 2D convolution with depth-separable convolution which will make the model lighter and faster; implement attention mechanisms to increase the accuracy of the network. The above modifications will allow achieving higher accuracy, resistance to overfitting and computational efficiency. Reproducibility aspects are also considered: saving the best weights, early stopping, data augmentations and metric evaluations of the results. In conclusion, the Dice and IoU metrics of the U-Net network with a set of the above modifications are presented, showing the impact of various changes in the architecture on the network. Based on the results of the work, the proposed modifications have shown their performance and efficiency and can be further used in solving classification and segmentation problems on rock samples.