Application of deep learning diagnosis for multiple traits sorting in peach fruit

Fruit quality is determined by multiple complex traits, which are difficult to diagnose by simple criteria and often require expert skills with a long experience. Nevertheless, current fruit sorting systems need a non-destructive, costless, and more rapid evaluation of fruit qualities. For peach, although many techniques have been developed for diagnosing fruit (internal) traits that determine commercial values, those techniques often require special facilities with high costs or take a long time for an assessment. Our study aimed to apply deep learning tech-nology to evaluate multiple peach fruit traits using only simple RGB images for practical applications. We tar-geted seven fruit traits fundamentally involving commercial fruit quality; skin color, flesh firmness, sugar content, and four internal disorders, including colorless early softening, split-pit, watercore, and damage from peach fruit moth. We performed binary classification and regression analysis for these traits by convolutional neural networks (CNNs). Binary classification is performed to judge whether a fruit trait exceeds a threshold or not for a given image. Regression analysis is performed to estimate the degree of a trait quantitatively. Their results suggested that CNNs can successfully diagnose multiple fruit traits and predict quantitative values from RGB images. We also applied an explainable AI (X-AI) technique to spot the hypothetical symptoms for each trait on a fruit image, giving novel interpretations for physiological reactions associated with each fruit trait.