Abstract: Linear-trajectory computed tomography offers significant potential in industrial online detection and plate-like structure imaging owing to its simple system structure and high scanning efficiency. However, its limited scanning direction and angular coverage result in insufficient depth resolution. Consequently, structures at different layers overlap in the projections, thereby causing interlayer aliasing and rendering it difficult to achieve high-quality results using conventional analytical and iterative reconstruction methods. Thus, this study proposes a deep reconstruction network based on the shift-and-add model. Under the geometric constraints of shift-and-add imaging, translation compensation between multi-view projections is parameterized, and the additive process is reconstructed in a network-based manner. This enables data-driven modeling to directly participate in multi-view projection fusion, thus enhancing the interlayer aliasing suppression capability while maintaining the physical consistency of imaging. Experimental results on the linear-trajectory PCB dataset indicate that the proposed method outperforms various comparison methods in terms of structural consistency and reconstruction quality. In particular, it demonstrates more stable performance in complex structures and small defect regions.