Enhancing optical non-destructive methods for food quality and safety assessments with machine learning techniques: A survey

Food quality and safety are critical to global health and economic stability, but traditional assessment methods, such as chemical assays and microbial culturing, are often destructive, time-consuming, and unsuitable for realtime and high-throughput applications. Optical non-destructive techniques, including imaging methods (e.g., red-green-blue (RGB) imaging, hyperspectral imaging (HSI)) and spectral methods (e.g., near-infrared (NIR) spectroscopy), offer real-time, precise, and non-invasive assessments while preserving sample integrity. However, the complex datasets generated by these techniques require advanced machine learning (ML) models for effective analysis. These methods generate complex, multidimensional datasets that align with ML approaches, unlocking advanced capabilities in data interpretation and decision-making. By integrating optical nondestructive techniques with ML models, ranging from classical algorithms like random forests (RF) and support vector machines (SVM) to deep learning architectures such as convolutional neural networks (CNNs), notable progress has been achieved in automating feature extraction, classification, and prediction tasks. This integration enhances the precision, scalability, and applicability of food quality and safety assessments, enabling tasks such as real-time grading, sorting, and microbial detection in diverse food systems. Advanced models like YOLO further expand the potential for real-time object detection in dynamic settings such as smart farms and food processing lines. Despite these advances, challenges remain in addressing the variability of food matrices, real-time processing limitations, and the need to integrate data from multiple optical models. This survey explores the integration of ML with optical non-destructive methods to enhance food quality and safety assessments, highlighting recent advancements and future opportunities.