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

Area of Science:

• Food Science and Technology
• Artificial Intelligence in Manufacturing
• Materials Science

Background:

• Three-dimensional (3D) food printing offers precise fabrication of plant-based meat analogues (PMAs).
• In-process quality control for 3D printed PMAs, particularly predicting protein composition and texture, remains a significant challenge.
• Non-destructive methods are needed for real-time monitoring during the 3D printing process.

Purpose of the Study:

• To develop a deep learning framework for non-destructive, in-process prediction of protein composition (wheat gluten, soy protein isolate, rice protein) and hardness in 3D printed PMAs.
• To utilize RGB surface images captured during printing for quality assessment.
• To establish a low-cost, real-time quality monitoring system for 3D food printing.

Main Methods:

• Developed a multi-attribute deep learning framework integrating semantic segmentation and sequence modeling.
• Employed DeepLabv3+ with a ResNet-50 backbone for segmenting printability features (pores, aggregation, spreading) from RGB images.
• Utilized an EfficientNet-Bidirectional Long Short-Term Memory (BiLSTM) model for multi-output prediction of protein content and hardness using image data and segmentation masks.
• Performed fivefold cross-validation for model performance evaluation.

Main Results:

• Semantic segmentation achieved a mean intersection over union (mIoU) of 0.558 and mean accuracy (mAcc) of 0.620 for structural features.
• The EfficientNet-BiLSTM model demonstrated high prediction accuracy: R² values of 0.857 (wheat gluten), 0.777 (soy protein isolate), 0.875 (rice protein), and 0.839 (hardness).
• Root mean squared error (RMSE) values were 4.035% (WG), 2.629% (SPI), 3.224% (RP), and 0.219 N (hardness), indicating reliable predictions.
• The framework successfully linked RGB image-derived morphology to compositional and textural attributes.

Conclusions:

• RGB image-based semantic segmentation effectively captures printing morphology relevant to PMA quality.
• The proposed deep learning framework enables simultaneous, non-destructive prediction of protein composition and hardness during 3D food printing.
• This approach offers a viable, low-cost solution for real-time quality monitoring in advanced food manufacturing systems.