The Research To Discriminate Wheat Quality Indicators Using Machine Vision Technology

Wheat is one of the three major reserve grains in China, and detecting wheat quality isvery importment to keep social stability and steady economic development. Determination ofwheat quality indicators currently are mostly based on the national standard and completed bysensory testing. Discriminating wheat quality by Machine Vision Technology can solve manyinsufficients which exist in manual inspection such as subjectivity, time consuming and laborintensive. In this paper, digital image processing, image analysis and pattern recognitionmethods were used, and the basic theory and methods used in Machine Vision Technology todetect wheat quality indicators like unsound kernels, impurities, moisture content and testweight were studied. The main contents and conclusions are as follows:(1)Images of unsound kernels, organic impurities, inorganic impurities, differentmoisture content of wheat were captured as well as different test weight of full grain kernelsand their cross-section images, then some image processings like gray-scale transformation,image binary, median filtering, image segmentation and morphological operations wereperformed to enhance image and dispel noises to ensure the images can meet the requirementsof image feature extraction.(2)After image pre-processing, some algorithms were used to extract morphological,color and textural features from different kernel images. Based on the analysis of parametersof three characteristic categories, we found that the parameters were different between soundkernels and unsound kernels as well as impurities, and these differences also exist in kernelsof different moisture content and test weight, and some characteristic parameters weresignificantly different. The principal component analysis and stepwise discriminant analysiswere used to select image features, and the principal components vectors or selection featureparameters were extracted as the inputs of pattern recognition to establish the appropriateidentification and detection model.(3)A three-layer back propagation neural network model with a structure of9-21-5wasestablished for identifying unsound kernels of wheat. The results showed that the recognitionrate of sound kernel, broken kernel, spotted kernel, sprouted kernel and injured kernel was93%,98%,100%,90%and85%, respectively, and the overall recognition rate was93%.Since the characteristic parameters of organic impurities, inorganic impurities and soundkernels were obviously variant, the single characteristic to its threshold based on decision treewere used, which turned out to be effective, and the parameters which had greatest contribution for identifying could be screened out by stepwise discriminant analysis. Whenusing the single characteristic threshold value method for all types of impurity identification,the recognition rate were90.0%~99.3%. When using a linear-function parametric classifier torecognize the sound kernels and organic impurities, inorganic impurities, all of the recognitionrate were100%.(4)For the detection of wheat moisture content, a back propagation neural network modelwith a structure of7-15-1was established, in which the coefficient between the estimatedvalue and the actual value was0.9546. For test weight identifying, wheat of different testweight were used and full grain images were captured as well as grain cross-section images,and linear-function parametric statistical classifier and back propagation neural networkmodel were established for identifying different test weight of wheat. The results showed thatthe features of full grain images could better reflect the difference of different test weight thanthe features of grain cross-section images. The overall recognition rate of over95%wasachieved for full grain images for the statistical classifier and BPNN model.(5)With the correlation algorithm results of this study, an analysis system of unsoundkernels of wheat based on Matlab software platform was preliminary developed. It isconcluded that this system is not only an effective way to identy unsound kernels of wheat,but also a reference for wheat quality testing based on machine vision.