Study On On-line Sorting System Of Citrus Color Based On Machine Vision

Citrus is one of the main cultivated fruits in Hunan,and its external quality needs to be sorted before entering the market to improve its commercial value.Among them,color selection is an important step in the detection and grading of citrus external quality.Given the problems of missing and repeated information collection in the application of machine vision in citrus color grading,this study developed an online citrus color detection and sorting system based on machine vision technology.The focus is on obtaining the full surface information of citrus without missing or repeating samples.Online classification of citrus color indicators is carried out through image processing algorithms and convolutional neural networks,and experiments are conducted to verify the classification accuracy and stability of the sorting system.The specific research content of the paper is as follows:(1)We have designed and built an online detection and sorting system for citrus color.We have completed the design of an image acquisition system on the existing citrus conveying production line,which mainly includes the design of the acquisition box shape,the selection of industrial cameras,lenses,and light sources,and the selection of light source lighting methods.A human-machine interactive sorting software for citrus color quality was developed using the VS2019.net platform,which achieved the detection of citrus color quality using traditional image processing algorithms and convolutional neural networks.(2)We studied an algorithm for extracting citrus coloring rate based on image processing.A citrus full surface image acquisition method was proposed to address the issues of information leakage and duplication in existing fruit full surface image acquisition methods.The method was used to uniformly and repeatedly capture images from different angles of the same citrus,and the feasibility of the method was verified through experiments.In the RGB color space,the contrast between citrus and background regions is improved by subtracting the B-channel grayscale image from the R-channel grayscale image,and the citrus and background regions are segmented using the bimodal threshold method.In the HSV space,citrus yellow regions are extracted based on the H component,and the two-dimensional yellow proportion of citrus is calculated.The average of the dynamically extracted feature values of citrus is used as the coloring rate for hierarchical classification.(3)A color grading method based on convolutional neural networks was studied.Firstly,the citrus color grading standards used in this article were provided,and the citrus dataset was completed.Through the shallow convolutional neural network and the transfer learning network MobileNet-v2,the same data set is trained to classify the orange color.Comparing the two networks,the results show that the Mobilenet classification accuracy is 96%,which is higher than the shallow convolutional neural network constructed.(4)Conduct an online comprehensive detection experiment on citrus color.Experiments were conducted on the stability and accuracy of online detection of citrus color characteristics at different speeds.The results showed that the maximum detection error of citrus color was within ±6%at speeds below 6 fruit cups/s.An online color grading experiment was conducted on 120 citrus fruits.The grading accuracy of traditional image processing algorithms was 92.5%,while the grading accuracy of deep learning methods was 95.8%.The sorting capacity can reach 4 tons/hour.The requirement for online non-destructive,fast,and accurate detection of citrus color has been achieved.The designed online non-destructive testing system for citrus color features greatly improves the efficiency of postpartum grading and provides a reference for subsequent research.