Automatic Sorting System for Citrus with Fast Visual Recognition and Improved Multi-Criteria Clustering

This paper presents a fully automated fruit grading system designed for Chinese fruit farmers, utilizing industrial vision and clustering algorithm for rapid and precise sorting. This system consists of an image processing subsystem for dual vision detection of fruit appearance, a clustering subsystem optimized with an improved K-Means clustering algorithm, and a mechanical subsystem scheme of a two-degree-of-freedom linear module and a three-coordinate manipulator. This mechanical system efficiently handles loading, packing, and unloading small patches of fruits in multi-stations. The sorting process was based on visual detection and qualification of fruit shape, peel appearance, and weight, during which a fruit-shape-rating algorithm using Sobel edge detection was developed to precisely identify fruit contours and diameters. In the clustering grading process, a multi-level evaluation model, adhering to the NY/T-1190 standard, was established to assess fruit weight, shape, peel color, and defects. To optimize one-time boxing efficiency and overall profit of small-batch fruits, the enhanced K-Means algorithm employed dimensionless fruit parameters for clustering. Experiments showed that the system achieved a 97.4% accuracy in detecting fruit defects, a 5.2% improvement in edge detection, and an 89.9% grading accuracy, with a 10.3% increase in same-batch sorting efficiency for small batches of 72 fruits.