Research On Appearance Quality Evaluation Of Rice And Wheat Grains Based On Machine Vision Technology

Rice and wheat are the main food crops and important commodity and reserve grain in our country.In the production,circulation,and consumption of grain,the detection of appearance quality plays an essential role in grading products.The traditional low-efficiency and high-error food quality and safety detection technology can not meet the current needs.In recent years,new detection technology has developed rapidly and is constantly replacing traditional methods.Machine vision technology has been proved to have significant advantages in grain species identification,grain external characteristics detection,imperfect grain grading,etc.It is not only convenient and fast but also has high accuracy.This paper analyzed the advantages and disadvantages of different equipment in grain image acquisition,summarized the appropriate methods to acquire grain appearance images,and then used the traditional image analysis technology and deep learning algorithm to preprocess the grain image to solve the problem of grain segmentation and counting.The main findings are as follows:(1)The traditional image analysis technique had better preprocessing results under single background conditions.The segmentation accuracy of rice grains was higher than that of wheat,and the increase in the number of grains would increase the difficulty of segmentation.The error-detection rate and the false-negative rate were not more than 3.5%,and the running time was controlled within 1 second.The deep learning algorithm had a remarkable effect in processing complex background images.The Faster R-CNN model had the best outcome.The accuracy was as high as 91%,and the running time was less than 2 second.The results showed that the linear model's accuracy based on image feature parameters could reach 96%,and the accuracy of the deep learning model was as high as 99%.However,the traditional image method was superior to the deep learning method regarding time cost and training cost.(2)The minimum bounding rectangle method is the most accurate method in the study of grain morphological characteristics.The deviations of the measured mean values of grain length,width and aspect ratio were within 0.15 min 0.10 mm and 0.08 mm,respectively,and the standard error were within 0.03 mm.The results of three-dimensional detection of grain based on binocular vision showed that the mean deviation of grain height and mitochondrial volume were 0.10 mm and 1.00 mm3.It could be used as an effective three-dimensional morphological index detection method.(3)Grain shadowing images with different degrees of fullness were processed to calculate the ratio of the circumscribed rectangle area(AR)and distance ratio(DR).The AR value of filled particles was usually above 4,but the AR value of unfilled particles was generally below 4,the DR value of filled particles was above 1.8,and the DR value of unfilled particles was below 1.8.Unfilled grain was detected more accurately using support vector machines(SVM)than using fixed thresholds for AR and DR,where AR was used better than DR.The error-detection rate of Indica rice varieties was slightly lower than that of Japonica rice varieties,but the false-negative rate of Indica rice was marginally higher,and neither of them was not more than 5%in Indica rice varieties.The neural network was trained by 19 conventional rice varieties,and the recognition results were 2.08%higher than that of the drain method.In the segmentation of broken rice grains,the edge center mode proportion method(ECMP)was not affected by the grain type and had the same segmentation effect on the complex grain adhesion.The detection accuracy of Indica rice and Japonica rice increased with the decrease of grain number,and the accuracy of 200 grains decreased by 2.71%compared with 100 grains,and the overall accuracy was over 95%.The minimum bounding rectangle method was used to identify the head rice with the highest accuracy,and the false detection rate and the missing detection rate were not more than 5%.With the increase of grain number,the detection accuracy of Indica rice and rice varieties had decreased,and the mean relative error was less than 3.21%.In the study of chalkiness recognition,the chalky part of rice grain was different from other parts of light transmission ability.The gray value of the chalky area was quite different from that of other parts,which can be used for accurate segmentation.But incomplete germ might affect the accuracy of recognition.By using SVM,the precise differentiation of the germ,back,abdomen,and heart can be realized.The accuracy of chalkiness of Indica rice and Japonica rice were 98.5%and 97.6%.(4)In the study of wheat appearance classification,the difference in the red value(r)?green(g)?blue(b)?and the color of grain(w),hue(Hu),saturation(Sd),brightness(V),length(L),width(W),height(H),volume(Vol),perimeter(P),area(S),eccentricity(C),angular second moment(ASM),correlation(COR),entropy(ENT),contrast(CON),the first moment(u1),the second moment(u2),the third moment(u3),etc.21 image characteristics of different grains were analyzed to determinate the appropriate preliminary parameters firstly.Furtherly,using scatter diagrams to confirm the best parameters and taking the best features and full parameters as model input parameters.Then,the mode accuracy of 7 machine learning algorithms was compared.It turned out,that there were great differences between broken grain and typical grain in texture and morphological characteristics.The optimal feature parameters were u1,u2,C,L,H,Vol.Logistic model constructed was considered as the optimal classification model.The accuracy was 100%,the false-negative rate was 0%,and all the AUC values were 1.There were great differences between dried and typical grains in texture and morphological features.The optimal feature parameters were ASM,COR,ENT,CON,S,W,H and Vol.The optimal classification model,Logistic,was built by a single parameter W.The model's accuracy was 100%,the false-negative rate was 0%,and all the AUC values were 1.There were significant differences in color,texture,and morphological characteristics between impurities and ordinary particles.In this case,the optimal characteristic parameters were w,Sa,u2 and C.The optimal classification model,Logistic,was built by a single parameter w;the accuracy was 98%and 100%,the false-negative rates were 2%and 0%,and all the AUC values were 1.When four varieties of grains were mixed,the optimal classification model,SVM,was built by full parameters.The accuracy of dried,impurity,normal and broken kernels were 100%,98%,98%,94%,the false-negative rates were 0%,2%,2%and 6%,the recall rates were 96%,98%,98%,88%,the false alarm rates were 4%,2%,2%,2%,AUC values were 1,1,1,0.99.All these showed that the model is reliable.