Fruit grading system by reconstructed 3D hyperspectral full-surface images

Abstract

Conducting hyperspectral imaging of fruit' entire surfaces while simultaneously evaluating their physical properties, such as volume and mass, can provide a richer dataset for comprehensive fruit classification. For this purpose, a fruit grading system using 3D hyperspectral full-surface images is developed, which is based on multi-view imaging of mirrors in the hardware structure design and relies on the virtual volume intersection (VI) algorithm and texture technology in software design. During the design process, a mathematical model for the mirror layout and system geometry parameters is established to determine the system's layout for scanning the entire surface of the sample. In practical applications, a prototype with 28 channels ranging from 400 to 1000 nm is developed for pear samples based on a sub-component control system, a spherical-cap cavity with flashing multi-color LEDs, and multiple side mirrors. The results obtained from this prototype reveal that the predicted volume (R2=96.18%) and mass (R2=98.18%) exhibit a high correlation with measured results and the hyperspectral data between bruised and normal pears is a significant difference. A dataset of pears of three qualities (large, small, and bruised) is prepared for comprehensive classification and resulting in an effective grading (95.33%), which shows that the proposed system is a potential solution for comprehensive fruit quality classification.