GReaT: Gated Re-Aggregation Transformers for Fine-Grained Visual Recognition

Plant disease recognition from leaf images is a highimpact visual classification problem that demands sensitivity to fine-grained, spatially localized symptoms while remaining robust to background variation and class imbalance. Convolutional neural networks offer strong locality bias but rely on fixed receptive fields, whereas Vision Transformers provide global token interaction yet often lack explicit mechanisms for periodic spatial consolidation on the patch grid. This paper introduces Gated ReAggregation Transformer (GReaT), a lightweight and additive augmentation for patch-token Transformers that explicitly reaggregates spatial evidence among patch tokens and reinjects the refinement through a learnable gated residual pathway. GReaT is implemented as sparsely inserted Gated Re-Aggregation Blocks (GRBs) that operate only on patch tokens, leaving prefix tokens and the original attention blocks unchanged, enabling stable warm-start training and straightforward integration into existing Transformer backbones. On the PlantVillage benchmark with 38 classes, GReaT consistently improves a DeiT-Base baseline trained under the same protocol, increasing accuracy from 98.88% to 99.67% with uniformly strong per-class performance. Qualitative analyses including confusion matrices, attention rollout, and t-SNE embedding visualizations indicate reduced confusion among visually similar categories, more symptom-focused spatial responses, and more discriminative global representations. These results demonstrate that explicit, gated spatial reaggregation provides an effective complement to global selfattention for fine-grained recognition.