MSPoreU-net: A novel U-net architecture to exploit multi-scale pores in segmentation of digital rock images
MSPoreU-net is a convolutional neural network workflow for better results in performance, run time, the number of trainable parameters, and network predictions in rock image segmentation. To assess the performance of the proposed MSPoreU-net architecture, we have tested and evaluated it on three different sets of rock images including Tight-carbonate, Sandstone, and Carbonate.
Schematic view of MSPoreU-net architecture in below figure. In this model, the sequences of three convolutional layers in the U-net architecture are replaced with the MSPore blocks. Furthermore, instead of using plain skip connections, the proposed MSPoreSkip block sequence is used.
The required packages to use this python repository are: 'os','numpy', 'scipy', 'h5py', 'tensorflow', 'matplotlib', 'keras', 'skimage', 'cv2', and 'pandas'. I recommend to use Anaconda which has all these packages installed except cv2 and tensorflow of which you can easily install from pip.
Example: Comparation of different pore scales: (a). Tight-carbonate sample has fine-size pores, (b). Sandstone sample has medium-size pores and (c). Carbonate sample has multi-scale coarse-size pores.
Mohsen Abdolahzadeh Kondori University of Tehran Phone:+989150465172 Email: MohsenKondori@ut.ac.ir MohsenKondori@yahoo.com