The observation of non-cooperative spacecraft using monocular cameras is a common perception task in space missions. However, target identification and tracking is inadequate for the requirements of increasingly complex non-cooperative target perception tasks. It is necessary to establish a method that can quickly and accurately create a structural description of non-cooperative targets through sequence images. This work proposes a method for reconstructing “structural semantic models” of non-cooperative targets. The method starts with component identification of typical components of a non-cooperative spacecraft, followed by 3D bounding box detection of the spacecraft body thereby providing a structured description of the 3D space in which the spacecraft main body is located. On this basis, we design an attribution inference method to evaluate the attribution probability of each typical component in the structured space so as to construct a graph-based “structural semantic model”. Finally, to address the problem of structured space misalignment during continuous observation, we design a graph-matching-based structural semantic model alignment method for continuous frames, which fuses the structural semantic models of successive frames to form a spatio-temporally consistent spatial non-cooperative target “structural semantic model”.