I selected Vision Transformers for leaf disease classification because they "revolutionize leaf disease classification", suggesting superior performance over traditional CNNs.
I am using the 'beans‑dataset' which contains labeled examples of Angular Bean Spot, Bean Rust, and Healthy leaves as a realistic testbed.
I chose accuracy as the primary evaluation metric for the model, aligning with typical classification objectives.
I reviewed Microsoft’s Swin Transformer as a breakthrough visual model that replaces flat tokenisation with hierarchical processing and shifted windows.
Swin uses hierarchical processing across stages to cut the quadratic complexity of ViT on high‑resolution images.
It introduces shifted windows to enrich feature perspectives beyond fixed‑size windows.
The architecture supports flexible input padding for varied image sizes.
These design choices make Swin scalable across tasks such as classification, object detection, instance segmentation, and masked‑image modeling.