Deep learning for non-destructive quality assessment and intelligent grading of postharvest fruits: A comprehensive review

Abstract Postharvest fruit loss presents a critical global challenge. Although deep learning (DL) has revolutionized quality assessment, achieving high accuracy in classification and defect detection, existing reviews predominantly focus on isolated technical aspects, lacking a systematic whole-chain perspective. To bridge this gap, this paper establishes a comprehensive review framework for postharvest fruit quality assessment and intelligent grading, encompassing the entire process from non-destructive data acquisition to industrial deployment. Following a "task-architecture-data" thread, we systematically compare the information characteristics and application scenarios of eight non-destructive sensing modalities (e.g., red-green-blue (RGB), hyperspectral, near-infrared, and X-ray imaging) based on DL data requirements. Furthermore, this review evaluates the applicability of fundamental DL architectures and synthesizes their recent advancements across three core domains: integrated appearance and internal quality assessment, pest/disease early warning, and intelligent grading. Crucially, we delve into the key engineering challenges of transitioning from laboratory settings to industrial sorting lines, addressing domain shift adaptation, model lightweighting, multi-sensor integration, and flexible grasping. Finally, we identify four current bottlenecks—data scarcity, insufficient generalization, the engineering deployment gap, and model interpretability—and propose corresponding future developmental directions. This review aims to foster the deep integration of advanced sensing technologies with neural networks, ultimately enhancing the efficiency of the global fruit and vegetable supply chain and reducing food loss.