In the process of reservoir drilling and development, it is of great significance to accurately extract, identify and evaluate the fractures in the formation to guide the drilling and development of oil and gas exploration. To solve the problem of imprecise fracture region segmentation by traditional methods, a fracture segmentation method based on Formation Micro-Scanner Image based on deep learning is proposed. Firstly, F-Criminisi algorithm is used to repair the blank strip with missing pixel information in the original FMI logging image. Then, a generative adversus-network based on U-Net is constructed, and dual attention mechanism is introduced to construct a fracture segmentation model to achieve accurate fracture segmentation under complex background. Combining pixel and edge information, loss function is designed to enable the model to more accurately segment the fracture and background region in the logging image and make the fracture boundary in the segmentation result clearer. In this paper, the proposed model is tested by using real FMI logging image of carbonate reservoir. The results show that the Dice coefficient of the proposed fracture segmentation method is 5% higher than that of the classical fracture segmentation model U-Net. This method can accurately extract fracture information from FMI logging images, and provides a basis for subsequent quantitative calculation of fracture parameters and logging interpretation, and has good practicability.