CNN–SVM hybrid model for varietal classification of wheat based on bulk samples

• Published in: European food research & technology Vol. 248; no. 8; pp. 2043 - 2052
• Main Authors: Unlersen, Muhammed Fahri, Sonmez, Mesut Ersin, Aslan, Muhammet Fatih, Demir, Bedrettin, Aydin, Nevzat, Sabanci, Kadir, Ropelewska, Ewa
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.08.2022; Springer Nature B.V
• Subjects: Agriculture; Analytical Chemistry; Artificial neural networks; Biotechnology; Chemical properties; Chemistry; Chemistry and Materials Science; Classification; Feature extraction; Food industry; Food Science; Forestry; Image resolution; Kernel functions; Neural networks; Original Paper; Support vector machines; Wheat; Deep learning; Support vector machine; Convolutional neural networks; Wheat classification
• ISSN: 1438-2377; 1438-2385
• DOI: 10.1007/s00217-022-04029-4

Determining the variety of wheat is important to know the physical and chemical properties which may be useful in grain processing. It also affects the price of wheat in the food industry. In this study, a Convolutional Neural Network (CNN)-based model was proposed to determine wheat varieties. Images of four different piles of wheat, two of which were the bread and the remaining durum wheat, were taken and image pre-processing techniques were applied. Small-sized images were cropped from high-resolution images, followed by data augmentation. Then, deep features were extracted from the obtained images using pre-trained seven different CNN models (AlexNet, ResNet18, ResNet50, ResNet101, Inceptionv3, DenseNet201, and Inceptionresnetv2). Support Vector Machines (SVM) classifier was used to classify deep features. The classification accuracies obtained by classification with various kernel functions such as Linear, Quadratic, Cubic and Gaussian were compared. The highest wheat classification accuracy was achieved with the deep features extracted with the Densenet201 model. In the classification made with the Cubic kernel function of SVM, the accuracy value was 98.1%.