부산대학교 도서관

Four deep convolutional neural network (DCNN) models, including Inception‐V3, ResNet50, VGG‐16, and a custom model were proposed to perform almond classification in five categories based on the adulteration of whole almond kernels with 0, 5, 15, 25, and 35% of broken almonds. Their performances were compared by statistical measures, including recall, specificity, precision, accuracy, and F1‐score. The results revealed that the models demonstrated great classification accuracies with the highest value achieved by ResNet50 (99%) and the minimum by the custom model (92.2%) while classifying different categories of almonds. However, compared to the three established models, the custom model had significantly lower precision and specificity values (minimum values 77.6 and 81%, respectively). Though, Inception‐V3, ResNet50, and VGG‐16 showed higher promises in overall almond classification compared to the custom model, their structures were bulker and more complex and required higher processing power and training time. The models attained good training and validation accuracy curves, showing a good fitting of the models for almond classification. The loss curves of the custom model were slightly underfit. Practical Application: California is the largest producer and supplier of almonds in the world. Despite technological advances, the grading of almond kernels is still a labor intensive and tedious process in California, let alone rest of the almond producers worldwide. The grading of almond kernels through machine learning algorithm using image processing and DCNN in this study can be extremely valuable for a rapid and automated grading of almond kernels on industrial scale. [ABSTRACT FROM AUTHOR]