The Research Of Wheat Quality Distinction System By Machine Vision

The objective of this paper is to use Machine Vision Systems (MVS) to inspect the grain quality, including detecting good and bad kernels, identifying the production year and distinguishing budded kernels and moldy kernels from normal ones.In MVS, the first step is to acquire digital images of the samples, and then the watershed transformation is used to segment wheat kernels, especially touching ones, from backgrounds. After extracting the morphological and color features of kernels, a neural network model as a classifier is to be constructed for pattern recognition.In detecting good and bad kernels, the system provides a classification accuracy of 90% above for good kernels and 80% above for bad kernels, which has been proved by other test samples. In identifying the production year, if kernels were stored carefully, the external features of them are almost the same, so computer could not make a right judgment. In distinguishing budded kernels and moldy kernels from normal ones, the accuracy for budded kernels reaches 80%, but the accuracies of normal and moldy kernels aren't as good as expected, maybe because washed moldy kernels disturbed recognition. In addition, the influences of morphological parameters and color parameters on the model are discussed.In the end, by selecting features through a combination of neural network and genetic algorithms, two GA-BP models are constructed, whose mean square errors are 0.001 and 0.00001. The results show the latter is better for a new BP model that can replace original one. Besides, the relations among all features are demonstrated by calculating the cosine value.