Study On The Quality Grading Method Of Shatangju Based On Computer Vision

In view of the problems such as low detection efficiency,high cost,large subjective factors,low accuracy due to standard disagreement,and less on-line continuous automatic classification research of the Shatangju,aiming at the difficulties of the classification of the Shatangju,the single camera and ring LED light source are used to extract the front image of the Shatangju,Through a large number of experiments and statistical analysis on the images with different freshness,the collection method,preprocessing method and segmentation method of the images are determined,and the quality related feature information such as fruit diameter,texture and surface defects are extracted from the images.The MLP model is designed and trained,and the traditional feature extraction,calculation and MLP are used to identify the hierarchical model Type a fusion was used to detect and grade the Shatangju.The main work and conclusions are as follows:(1)According to the characteristics of Shatangju,a recognition system is built with light source camera and other hardware to confirm the image acquisition method and complete camera calibration,transform RGB and HSI color models and compare the characteristics of satsuma orange under each color model,image preprocessing and so on,make sufficient preparation for the following smooth feature extraction.(2)The advantages and disadvantages of image edge segmentation methods based on gray level similarity and gray level discontinuity are compared.Otsu threshold is proposed to separate background of orange;compared with the methods of edge extraction,the effect pictures of the first-order and second-order differential edge operators which are commonly used under different lighting methods and two light source combinations are adjusted,and it is concluded that lighting methods and light source selection have great influence on the edge operators;and verify and compare the different extraction methods of the fruit diameter of several kinds of Shatangju,find the most suitable calculation method of the fruit diameter fitting circle,in addition,compare the calculation methods of the area circle ratio of two kinds of Shatangju,find a more accurate calculation method of the ratio of the area and the circumference,and determine the calculation for the extraction of the fruit diameter and the circle ratio of the appearance quality grade characteristics of the subsequent Shatangju Scheme.(3)The fruit diameter and shape of the Shatangju are calculated by the image of the Shatangju,which is separated from the background.After the analysis and statistics of the gray values of the RGB and HSI images with different freshness,the image segmented between the fixed threshold 25-128 of the s channel and the image of the H channel are used for mask operation,Then,the H channel was segmented by the threshold of 28 °-72 °,and the fruit stalk was extracted successfully.The fruit stalk was used to identify the freshness of the Shatangju.(4)In H channel,the region of fruit stalk was removed,and then the coverage of cyan pixels in the whole peel was calculated by threshold segmentation of cyan pixels,and the maturity of the Shatangju was determined.(5)The method of gray level co-occurrence matrix is successfully used to extract texture features.Six texture features of the gray level image of the Shatangju are used as the basis to determine whether the peel of the Shatangju is smooth,and the accuracy of the recognition results is calculated.Finally,it is found that the linear segmentation with simple features can not meet the requirements of the recognition accuracy of the whole system.the recognition rate of the commonly used method of calculating the sum of the number of regions and the R,G and B components of the comparative region combined with the metric range is not ideal,Instead,the fast Fourier transform image is used to identify the cracked fruit in the frequency domain,and then the cracked fruit and the scarred fruit are distinguished by morphology.The accuracy of the latter can completely meet the needs of system classification.(6)Compared with the national standard quality level list of Shatangju,this paper developed the level recognition setting decision table suitable for this research topic,and designed the overall framework flow of the whole quality recognition system,in which 6parameters of texture characteristics were used to construct 6-8-2 BP network structure for learning and training a network recognition model of skin smoothness,and the accuracy of the model classification reached 97%,It can meet the requirements of grading the smoothness of Shatangju peel;Firstly,the fast Fourier transform is used to determine the peel fracture and scar,then the traditional visual feature judgment method and MLP are used to determine the smoothness of the peel to form a discrimination tree,and finally the final grade result of the Shatangju is obtained by decision fusion judgment.(7)Integrate all the feature recognition algorithms,generate a GUI program system,use the system verification samples to get the decision result pictures,and analyze the overall system recognition rate of 90%.Through the comparison of experimental data,the results show that the system has a certain practicality.The above research results provide technical support and theoretical basis for the automatic and scientific determination of the appearance quality of Shatangju.It has a certain theoretical and practical value for promoting the classification and screening of the Shatangju and promoting the trade development of the Shatangju.