Online detection and sorting of passion fruit quality based on machine vision using multi-indicator

Abstract

[Objective] Refining the accuracy and intelligence of passion fruit quality assessment. [Methods] The study used the capabilities of OpenCV along with a compact neural network architecture, MobileNetV3_large_ssld, to accurately determine the fruit's diameter, ripeness, and the degree of its wrinkling. The diameter measurement was achieved by analyzing the short side of the fruit's minimum bounding rectangle. The ripeness assessment was based on the pixel ratio of H component values within specific ranges （H∈[0,10]∪[156,180], [11,25], [26,34], [125,155]） in the HSV color space. Furthermore, the study developed a MobileNetV3_large_ssld model to evaluate the wrinkling of the fruit's skin. Leveraging these three key indicators, a comprehensive fruit quality evaluation model was established using a rating scale approach, and an online detection and sorting system was subsequently developed. This system employed KNN background subtraction to extract the fruits target, excludes stems, and used interval frame sampling method to capture single image for each fruit from the video. The comprehensive evaluation model was utilized to assess the quality of passion fruits, which were then sorted into their appropriate grade channels through a sorting mechanism. [Results] The test results indicated a high degree of consistency between the system's sorting and manual sorting, with an overall accuracy of 97.02%. The consistency rates for top-grade fruits, first-grade, and second-grade fruits were 95.51%, 97.84%, and 100%, respectively. [Conclusion] This system could be used for online detection and sorting of passion in different quality grades.

Keywords

• postharvest handling
• passion fruit
• image processing
• object recognition
• neural network
• sorting