Fine-Grained Soybean Variety Recognition Using Swin Transformer with Differential Attention and Dynamic Channel Aggregation

Soybean variety recognition supports germplasm management and intelligent agricultural inspection, but remains challenging because varieties often show subtle inter-class differences, large intra-class imaging variation, and imbalanced samples. This study adopts a cascaded pipeline that combines YOLO-based object localization, manual verification and cropping, and fine-grained classification, allowing the classifier to learn from standardized cropped soybean seed images rather than original whole images. To improve Swin Transformer for this task, we propose Swin-Diff-DCA, which introduces Dynamic Channel Aggregation (DCA) in Stage 3 for middle-level local feature reuse and a differential attention branch in Stage-4 window attention for deep discriminative enhancement. On a dataset containing 25 soybean varieties and 1511 test images, Swin-Diff-DCA achieves average results of 87.18% Accuracy, 82.33% Macro-F1, and 87.44% Weighted-F1 across three random seeds under the current split and evaluation protocol. It outperforms RegNet, ResNet-50, ViT, the Swin Transformer baseline, and single-module variants. The results show that combining middle-level feature reuse with deep differential enhancement improves cropped soybean variety classification, while low-sample and visually similar classes remain the main sources of misclassification.