Grain classification using hierarchical clustering and self-adaptive neural network

Self-adaptive back propagation neural network (BPNN) models based on hierarchical clustering were developed to classify corn kernels. To generate the sample sets, randomly selected kernels were divided into seven classes using multiple clusters, including three classes of flat kernels, three classes of round kernels and abnormal class. Further, the stepwise discriminant analysis was conducted to select eleven morphological features. For each model, features with small discriminatory power were removed respectively. Finally, several self-adaptive BPNN classifiers were combined for corn kernels classification. To construct the self-adaptive BPNN classifier, not only the resilient backpropagation algorithm was used to train neural network, but Nguyen-Windrow was used to generate initial weight and bias values. Therefore, the convergence of self-adaptive BPNN was accelerated, and the problem of staying at the local minimum points was overcomed. Experiments showed that, kernels of every class had good uniformity in morphology after clustering and average accuracies of the whole network were over 90%.