The invention provides a multi-decoder full convolutional neural network and a corresponding microscopic structure identification method thereof, the multi-decoder full convolutional neural network isused for semantic segmentation of ceramic matrix composite material perform XCT slices, and an initialization method, a class balance method, a training method and the like of a network structure aredetermined. According to the method, the problem of edge blur or content information loss in the feature map caused by different decoders is solved, the accuracy of microscomic structure recognitionis improved by combining deep learning and semantic segmentation tasks, and the robustness of an algorithm is enhanced. The multi-decoder full convolutional neural network provided by the invention can be used for ceramic matrix composite preforms of different weaving types, including 2.5-dimensional weaving, three-dimensional four-direction weaving structures and other all types, the applicationrange is wide, and identified microscopic structures comprise fiber bundles (warp yarns and weft yarns), holes and matrixes.