Multi-attribute prediction of protein composition and hardness using semantic segmentation for 3D-Printed plant-based...

Three-dimensional (3D) food printing enables layer-by-layer fabrication of plant-based meat analogues (PMAs), but reliable in-process quality evaluation remains challenging. This study developed a multi-attribute deep learning framework for the non-destructive, in-process prediction of formulation-derived protein composition and hardness from RGB surface images captured during printing. The target attributes were wheat gluten (WG), soy protein isolate (SPI), rice protein (RP), and hardness. A DeepLabv3+ model with a ResNet-50 backbone was used to segment printability-related structural features, including pores, aggregation, and spreading, achieving a mean intersection over union (mIoU) of 0.558 and mean accuracy (mAcc) of 0.620. RGB images and segmentation masks were then used as inputs to an EfficientNet-Bidirectional Long Short-Term Memory model for multi-output prediction. Fivefold cross-validation showed R² values of 0.857 for WG, 0.777 for SPI, 0.875 for RP, and 0.839 for hardness, with corresponding root mean squared error (RMSE) values of 4.035%, 2.629%, 3.224%, and 0.219 N, respectively. These results demonstrate that RGB image-based semantic segmentation can capture printing-related morphology and support the simultaneous prediction of compositional and textural attributes. The proposed framework provides a low-cost, non-destructive approach for image-based quality monitoring in 3D food printing systems.