Abstract:In this paper, an improved PSPNet network is proposed to automatically identify fractures in electrical imaging logging images, which is difficult to extract fracture features and leads to low segmentation accuracy and large calculation of network parameters. Firstly, the backbone network in PSPNet is replaced with the optimized MobileNetV3 network, which can significantly reduce the number of network parameters and the amount of computation. Secondly, the Asymptotic Feature Pyramid Network (AFPN) is introduced to increase the interaction of multi-scale information and enhance the recognition ability of small cracks. Then, Multi-Depthwise Conv head Transposed Attention (MDTA) was introduced to extract global features and improve the extraction ability of key information. Finally, the combination of Focal Loss and Dice Loss was used as a loss function to solve the problem of unbalanced proportion of data sets. The experimental results show that the improved PSPNet network has a good segmentation effect on the fracture in the electrical imaging logging. Compared with the PSPNet network, mIoU improved by 3.17% and mPA improved by 6.38%. In addition, the number of parameters, calculation amount and weight of the proposed algorithm are reduced by 94.3%, 95.7% and 93.8% respectively compared with the original model. At the same time, the crack identification system based on CIFLog is developed, which can meet the practical needs of the electrical imaging logging.