On-line detection of orange soluble solid content using visible and near infrared transmission measurements

A prototype of on-line system developed by ourselves was used to non-destructively inspect orange quality. This system includes three main parts: machine vision part for fruit external quality detection, visible and near infrared (Vis-NIR) spectroscopy part for fruit internal quality detection, and weighing part for fruit weight detection. Fruit scrolled on the roller in the machine vision part, while stopped scrolling before entering the Vis-NIR spectroscopy part. Therefore, fruit positions and directions were inconsistent for spectra acquisition. This paper was aimed to study the influence of fruit detection orientation on spectra variation and model estimation performance using the on-line system. The system was configured to operate at typical grader speeds (0.27m/s or approximately three fruit per second) and detect the light transmitted through oranges. Stepwise multi linear regression models were developed for fruit with consistent directions and inconsistent directions in the wavelength range of 600-950 nm, and gave reasonable calibration correlations R2=0.89-0.92 and low cross validation errors (RMSECV=0.44-0.56%). The calibration model with spherical samples only turned out the best prediction results, which has lowest RMSEP of 0.56%-0.63% for different fruit orientations. It can be seen from the study that fruit shape would influence the fruit orientation for spectra aqcuiring of spherical samples after scrolling, and would further influence the modeling resutls. It is better to acquire spectra and establish models for sampels with different shapes separately and then applying them based on shape detection resutls to improve the soluble solid content (SSC) prediction accuracy.