GRAPE LEAF VARIETY RECOGNITION BASED ON THE AF-SWIN TRANSFORMER MODEL

Abstract

Aiming at the problem of differentiated cultivation strategies for different grape varieties, the AF-Swin Transformer model is proposed in this study. Firstly, Focal Loss is used to effectively tackle data imbalance in grape leaves. Secondly, the AdamW optimizer is selected to better control model complexity and improve generalization. The results show that the training accuracy of AF-Swin Transformer model is 7.87 percentage points higher than that of the original Swin Transformer model. Precision and recall improved by 4.4 and 4.3 percentage points, respectively. This study enables accurate automated variety monitoring within vineyard cultivation systems, assisting growers in implementing targeted cultivation strategies.