Harnessing deep learning for wheat variety classification: a convolutional neural network and transfer learning...

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Non-destructive food assessment methods, particularly computer vision, are increasingly important due to low computational costs.
• Existing wheat classification models often suffer from limited data and a narrow range of varieties.
• The need for robust models capable of classifying a wide diversity of wheat varieties is critical.

Purpose of the Study:

• To assess the applicability of convolutional neural network (CNN) models for classifying a large number of wheat varieties.
• To develop and evaluate a novel, four-layered CNN model from scratch.
• To compare the performance of the novel CNN model against popular pretrained architectures using transfer learning.

Main Methods:

• Preparation of multi-view images for 124 distinct wheat varieties.
• Development of a custom four-layered convolutional neural network (CNN) model.
• Application of transfer learning to train established architectures: DenseNet201, MobileNet, and InceptionV3.

Main Results:

• The proposed CNN model achieved a classification accuracy of 95.40%, outperforming DenseNet201 (92.41%), MobileNet (90.54%), and InceptionV3 (83.47%).
• The custom CNN model demonstrated superior performance despite high computational demands, indicating its effectiveness.
• The use of a multi-view, large-image dataset was crucial for achieving high accuracy in classifying numerous wheat varieties.

Conclusions:

• The developed CNN model shows significant promise for accurate, non-destructive wheat variety classification.
• Further fine-tuning of hyperparameters and evaluation of other models are recommended to enhance accuracy.
• The study will release its image datasets to facilitate further research in wheat classification methodologies.