Comparison of two neural network architectures for classification of singulated cereal grains

摘要

A digital image analysis algorithm was developed to facilitate classification of individual cereal grain kernels (barley, Canada Western Amber Durum (CWAD) wheat, Canada Western Red Spring (CWRS) wheat, oats, and rye). A total of 230 features (51 morphological, 123 color, and 56 textural) were extracted from 7500 high resolution color images of each type of grain using the developed algorithm. A four-layer back-propagation neural network (BPN) and a specialist probabilistic neural network (SPNN) wereevaluated for classification accuracies. The BPN used a sigmoid scaling function for input nodes and sigmoid activation function for nodes in the hidden layers. Five different data sets were used for training, testing, and validation. The BPN based classifier outperformed the SPNN classifier for all grain types. Using various features models, the average classification accuracies for BPN were 96.4, 90.8, 98.0, 95.5, and 964% for barley, CWAD wheat, CWRS wheat, oats, and rye, respectively. For the SPNNclassifier, the average classification accuracies were, 91.5, 84.7, 95.3, 88.4, and 93.3% for barley, CWAD wheat, CWRS wheat, oats, and rye, respectively.