Development of an optimally designed real-time automatic citrus fruit grading–sorting​ machine leveraging computer vision-based adaptive deep learning model

Abstract

Conventional automation approaches for postharvest operations are plagued by time and data inefficiency seldom leading to suboptimal solutions. Automatic machines often require highly skilled software professionals for calibration and reconfiguration thus making the technology prone to high costs. Contemporary sensors and smart devices capable of handling deep learning image analytics have been employed in the present study for the development of an automatic machine that performs postharvest operations, like—washing, vision-based sorting and weight-based grading of citrus fruits with much reduced human effort while achieving excellent performance for the designated tasks. Accuracy of performance was ensured by the optimal design of mechanical components carried out by kinematic synthesis and dimensional analysis. The machine was equipped with an effective custom lightweight CNN model “SortNet” that was designed and tuned to carry out vision-based classification of citrus fruits into “accept” and “reject” based on surface characteristics. SortNet was less complex and took less computational time while exhibiting comparable accuracy with respect to existing state-of-the-art pre-trained deep learning models. An embedded system operated by a single-board computer was used in the weight grading section for segregating fruits based on three weight categories. Evaluation, realization and transferability of the above said strategy was demonstrated by the real hardware with physical actuators working in real-time to serve as proof-of-concept for a sustainable solution to postharvest automation of citrus fruits.