Lightweight deep learning framework for quality assessment of rolled fondant

Rodríguez E.; Pauccar G.; Carrizales J.; Inga Espinoza C.H. · LWT (2026) · DOI: 10.1016/j.lwt.2026.119580

Resumen

Inconsistent quality and operator subjectivity in rolled fondant production hinder process standardization. This study presents an integrated vision-based quality verification system for the nondestructive assessment of sheeted fondant, focusing on color class, spatial uniformity, and surface porosity. The framework combines a calibrated sRGB to CIELAB imaging pipeline with three custom ultra lightweight deep learning models (19k to 766k parameters). To ensure statistical rigor, performance was evaluated using 50 captures per model to establish 95 percent Confidence Intervals, and benchmarked against YOLOv5, YOLOv8, and YOLOv11 architectures on a Raspberry Pi5. The proposed models achieved superior metrological reliability. The classification models reached 100 percent accuracy and perfect uniformity estimation within the highly controlled radiometric dataset, overcoming the low prediction confidence and high false positive rates of the heavily parameterized YOLO networks. Furthermore, the custom fully convolutional network accurately quantified surface porosity (38.27 percent), avoiding above 95 percent saturation errors of the YOLO segmentation heads and demonstrating that continuous micro-textural defects require semantic rather than instance segmentation. While multi second inference times currently restrict deployment to final state processes rather than continuous high speed inline monitoring, these results validate a highly repeatable, parameter efficient, and low-cost metrological tool for smart food manufacturing environments. Copyright

Instituciones: Universidad Peruana de Ciencias Aplicadas