A Hybrid Model Approach Based on Swin Transformer and EfficientNetV2 for Maize Variety Classification

In this study, two different deep learning-based models were proposed for the classification of the maize varieties Chulpi Cancha, Indurata, and Rugosa. In the first stage, a single model was developed using the Swin Transformer architecture with an attention mechanism. This model was then integrated with EfficientNetV2 to create a hybrid structure. The developed models were tested on a dataset consisting of 1050 images with a fixed background and high resolution. The Swin Transformer model produced successful results with 99.37% accuracy, while the hybrid model achieved 100% test accuracy, accurately classifying all samples. The findings demonstrate that the Swin Transformer and EfficientNetV2-based hybrid architectures offer high discrimination power and generalization capacity in image-based classification of maize varieties. Future studies are recommended to conduct additional tests using images taken under different environmental conditions and larger datasets encompassing a wider range of varieties.