Integrated Intelligent Digital Rock Images Segmentation Based On U-Net Deep Neural Network
Jun Luo, Qiang Chen, Yinchun Tian, Guoqiang Wu, Haoyu Shi, Lin Liu
U-Net is a new type of the full convolution network (FCN) developed in 2015. The U-Net model is characterized by end-to-end structure, which allows that the input and the output are both images with the same size. This makes it possible for finer semantic segmentation at pixel-level. Moreover, the much less training set size and faster training are some of the other advantages of U-Net. Thus, U-Net has been widely used in image segmentation with good performance especially in the medical field. Based on the improved U-Net model and image processing algorithms, intelligent segmentation workflow for digital rock images was built. The workflow consists of five parts, namely image acquisition, image preprocessing, data preparation, model training and model application. In addition to two-class segmentation problems, the workflow is also applicable to multi-class segmentation. Based on this workflow, two-class and three-class segmentations were conducted on unconsolidated sandstone and carbonate CT images.
The results indicated that in the case of two-class segmentation, when utilizing a single unconsolidated sandstone CT image for training and testing, the training accuracy could be optimized above 90%. Computed connected porosity and digital rock permeability errors between U-Net segmented model and artificially segmented model were 4.36% and 9.12%. These results showed a relatively good performance. However, when the trained model was extended to a carbonate CT image which had never been trained, the performance dropped dramatically with most pores not recognized. By adding the same number of carbonate CT images as the unconsolidated sandstone to enhance the training set, as well as optimizing training parameters such as learning rate and training epochs, training accuracy could be improved to 95%.
Furthermore, the improved model was applied to segment other rock CT images from different oilfield with 2 carbonate samples and 2 unconsolidated sandstone samples. The results showed that the average computed connected porosity and digital rock permeability errors of carbonate samples between U-Net and artificially segmented models were 3.18% and 7.45%, respectively, compared with that of 2.41% and 6.67% of unconsolidated sandstone samples. The improved model indicates a significant improvement of segmentation accuracy and generalization performance compared to the first trained model.
In the case of three-class segmentation, namely the pore phase, the matrix phase and the pyrite phase, the U-Net model was trained with 60 images from a single sandstone CT sample. The accuracy of the training was 86%. Application of the trained three-class segmentation model in a new sandstone sample with 0.45% pyrite showed that for pore, matrix and pyrite proportions, the errors between U-Net and artificially segmented models were 4.86%, 1.99% and 6.67%, respectively. These results indicate an extraordinary segmentation accuracy especially for pyrite phase. Therefore, the U-Net model has an excellent performance in digital rock image segmentations for both two- and multi-classification and its use is highly recommended in digital rock physics technology.
