DETERMINING WHEAT VITREOUSNESS USING IMAGE PROCESSING AND A NEURAL NETWORK

Abstract

A realtime, imagebased graininspection instrument (GrainCheck 310) was used to develop backpropagationANN models to classify durum wheat kernels based on their vitreousness. Singlekernel images were created throughpreprocessing. HSI color features of image rows and columns were used as the inputs to the ANNs. Several 11class, 3class,and 2class ANN models were trained with different numbers of hiddenlayer nodes and training epochs. Classification ratesof 85% to 90% were achieved for the vitreous and nonvitreous classes. For all nonvitreous kernel subclasses, except thebleached subclass, the classification rates also reached 85%. A correction algorithm was developed to overcome thedifficulty in measuring mottled kernels caused by the orientation of kernels under the camera. A 2class ANN modeldeveloped in this study was tested on two GrainCheck 310 machines. The average difference between the classification resultsof these machines was 1.5%, indicating a good model transferability between machines. Comparisons also showed that theperformance of the machines is more consistent than human inspectors.