Abstract: To address the key challenges in three-dimensional (3D) mesostructure modeling of novel composite solid propellants, such as insufficient component segmentation accuracy, truncation and interlayer defects easily caused by the traditional method of planar segmentation followed by 3D reconstruction, and the difficulty in effectively distinguishing components with similar grayscale values, a mesoscopic finite element reconstruction method integrating μ-CT experiments and spatial digital image processing technology was proposed. This method adopts an optimized technical route of 3D reconstruction prior to spatial segmentation, which preserves the spatial topological relationships of mesoscopic components. Initially, two-dimensional slice images of the propellant specimen at different heights are acquired via μ-CT scanning, and preprocessed to eliminate scanning noise and invalid background regions. Subsequently, all preprocessed slices are stacked along the height direction to construct a three-dimensional grayscale array containing complete spatial position and grayscale information. On this basis, targeted segmentation strategies are implemented for different components respectively, in combination with the grayscale ranges, spatial morphological characteristics and size distributions of each component obtained from SEM/EDS tests. For initial defects, background interference is eliminated by combining the lower grayscale limit with connected domain volume screening. For lead (Pb) particles with independent, non-overlapping grayscale ranges, segmentation is completed directly via grayscale thresholding. For ammonium perchlorate (AP) and aluminum (Al) particles, sequential spatial morphological operations are performed on the basis of preliminary grayscale screening, coupled with screening by connected domain equivalent diameter and voxel count, to remove interference from burrs, noise and agglomerates. For HMX and the viscoelastic matrix with overlapping grayscale values, the mixture region is first obtained through preliminary grayscale range screening. The optimal segmentation threshold is then determined using the Otsu adaptive threshold method. Finally, based on the morphological characteristics and spatial dimensions of HMX, spatial connected domain analysis is applied to filter out undersized matrix noise and oversized agglomerates, realizing effective separation between the two components. The segmentation results show that the volume fractions of AP, Al, Pb, initial defects, HMX, and the viscoelastic matrix are 8.28%, 4.37%, 0.254%, 0.16%, 39.74%, and 47.20%, respectively, which are in good agreement with experimental measurements. Compared with conventional methods, the method proposed effectively avoids truncation and interlayer defects caused by planar segmentation, and more realistically reflects the 3D spatial distribution characteristics of each component. Furthermore, a 3D finite element model of the representative volume element (RVE) is established based on the segmentation results, which provides more reasonable structural input and reliable technical support for multiscale mechanical analysis, damage evolution simulation of solid propellants, and structural integrity assessment of propellant grains.