Online volume measurement of sweetpotatoes by A LiDAR-based machine vision system

Volume is an important phenotype and shape descriptor of sweetpotato storage roots for postharvest quality assessment and breeding programs. The standard water displacement method for measuring fruit volume is laborious, subject to human observation errors, and often destructive. Three-dimensional (3D) imaging-based machine vision provides an alternative for rapid and non-destructive volume estimation. However, existing machine vision-based methods usually require manually positioning samples or sensors to acquire different views of samples and hence are not suited for online, high-throughput applications. This study presents a novel machine vision system that comprises a custom-designed roller conveyor system to simultaneously transport and rotate sweetpotatoes for full-surface imaging, and a consumer-grade LiDAR (light detection and ranging) camera to acquire color and depth images. Sweetpotato storage roots of Beuregard were imaged online by the system with six consecutive scans obtained for each sample. Regression models were built for volume estimation using MLR (multiple linear regression), SNN (shallow neural network), and DNN (deep neural network) based on geometric features extracted from segmented fruit masks and the Alpha shape reconstructed from depth images. A feedforward four-hidden-layer DNN pre-trained by stack auto-encoders achieved the best accuracy of 97.9% (R2 = 0.993 and RMSE = 9.5 cm3), compared with the accuracies of 94.2% (R2 = 0.958 and RMSE = 27.2 cm3) and 96.2% (R2 = 0.976 and RMSE = 19.1 cm3) obtained by MLR and SNN, respectively. The developed 3D machine vision system is useful for online, rapid volume measurement of sweetpoatoes and can be potentially applied to other fruits.

Automated summary

This study presents a machine vision system for volume measurement of sweetpotatoes, using automated imaging and deep learning models to achieve accurate and non-destructive volume estimation.