Development of tomato non-destructive measurement system based on machine vision

Abstract

Tomato size can be used as an indicator for judging its quality and yield, which is helpful for quality assessment and yield management. In order to fulfill the task of fast and accurate measurement of tomato size, in this study a tomato size integrating segmentation and detection measurement system were proposed and constructed, which can complete the calculation of tomato size calibration coefficients, the recognition and segmentation of tomato images and the measurement of tomato size. An improved model for tomato recognition and segmentation based on machine vision is first presented in this study. The model constructs a lightweight network model Tomato-YOLOv8s-Seg based on YOLO v8s, improves and replaces all the convolutional layers with GHostconv, and redesigns the C2f_SE module, introduces the SE attention mechanism on top of the C2f module, and additionally fuses the shared parameters and the feature fusion module BiFPN on the detector head processing to reduce the reference numbers; secondly, a new tomato size measurement scheme is proposed, which firstly measures the actual size of the tomato using the head measurement system, then uses the proposed network model segmentation to get the number of pixels of the tomato size, and calculates the calibration coefficients by the calculation of both of them and then calculates the actual size of the other tomatoes in the image; finally, experiment are carried out by using the tomato measurement system constructed in this study. The experiment results show that the proposed lightweight network model deployed on OrangePi5pro had an accuracy of 92.1%, an FPS of 54, a MAPE of 1.60% and 4.79% for measuring tomato diameter and volume size, an RMSE of 0.09 cm and 6.51cm³, and a measurement time of 0.5 s per tomato. The system provides a fast, accurate, and low-cost solution for non-contact measurement of tomato size.