The homogeneity of the seeds is an important factor in terms of processing, transportation, storage, and product quality of agricultural products. It is possible to classify the grain polymorphism of barley cultivars, which are economically important among cereal crops, in a short time with computer vision methods with high accuracy rate and almost zero cost. In this research, a novel image database consisting of 2800 images were created to classify 14 barley cultivars. Six different deep convolutional neural network models were designed based on a transfer learning method with pretrained DenseNet-121, DenseNet-169, DenseNet-201, InceptionResNetV2, MobileNetV2 and Xception networks. The models were trained and evaluated with test-time augmentation method, the best performance was obtained from DenseNet-169 model with average 96.07% recall, 96.29% precision, 96.07% F1-score, and 96.07% accuracy on a test set independent of the training set. The results showed that the transfer learning method performed using additional layers such as dropout and data augmentation with sufficient data samples in these images with high similarities prevented overfitting by increasing the model performance. As a result, it can be suggested that the provided web tool based on the transfer model has an encouraging performance in identifying seedswith a high number of cultivars such as barley.
Read full abstract