Dense sandstone material decomposition based on improved convolutional neural network

Energy spectrum computed tomography can provide quantitative information of scanned objects and realize material decomposition. At present, the material decomposition method based on neural networks overcomes the limited decomposition effect of traditional iterative algorithms. However, the performance of traditional neural networks in feature detail recovery is still not satisfactory. To improve the material decomposition accuracy, a material decomposition method based on a Resnet and Squeeze excitation network (RS-Net) is proposed. The proposed method uses the structure of the U-Net network and Resnet-152 as the backbone network to extract multi-scale features. Parallel asymmetric convolution is used to complete the large kernel convolution, which reduces the number of parameters and computation of the network. The HD-SE attention mechanism is introduced in the decoder part to help the network recover the image features. Hybrid loss supervised network learning is used to improve the decomposition accuracy of the network. The feasibility of this method is verified on simulated rock and artificial sandstone datasets. The simulation and experimental results show that RS-Net combined with mixing loss can retain more internal details of the image, the decomposed image edge is clearer, and the image quality is higher.