Automated Segmentation of Manufacturing Defects in 3D ‐Printed Composite Materials Using an Optimized Deep Learning Approach

Precise segmentation of internal defects in composite materials is a crucial foundation to analyze their mechanical properties based on characterization. Because 3D printed composite materials have numerous and irregularly distributed voids, conventional segmentation methods suffer from drawbacks such as being time‐consuming or having low accuracy. This study proposes CAMPS‐UNet, a novel lightweight segmentation network for identifying voids in x‐ray computed tomography (X‐CT) images of 3D printed carbon fiber reinforced polymers (CFRPs). Built on a Vision Mamba UNet backbone, CAMPS‐UNet integrates Coordinate Attention (CA), multiscale Atrous Spatial Pyramid Pooling (ASPP), and a SegFix boundary refinement module. This structure jointly captures long‐range dependencies, multiscale context, and fine boundary details, enabling precise and efficient void segmentation. Ablation studies confirm that every module contributes to improving segmentation accuracy. CAMPS‐UNet consistently outperforms conventional architectures, achieving superior results across standard segmentation metrics: the highest Dice coefficient and MIoU. In 3D reconstruction, CAMPS‐UNet enables reliable quantification of the voids volume fraction, together with the corresponding size and shape distributions. Furthermore, CAMPS‐UNet demonstrates robust generalization capability, highlighting its effectiveness across different imaging conditions.