plant-pathology-2021

Dataset Overview

Dataset Name

• Plant-Pathology-sysu-2023

Dataset Purpose

• Used for leaf‑disease classification tasks in plant pathology.

Models Used

• ResNet50, VGG16, SwinTransformer, EfficientNet, MobileNetV2
• SwinTransformer_STN (an improved SwinTransformer that incorporates an STN module)
• Attention‑SwinTransformer (a multi‑model‑based design)

Research Highlights

• Introduced a Spatial Transformer Network (STN) into the SwinTransformer model to increase sensitivity to spatial transformations in leaf‑disease images.
• Combined multiple models and added an attention module to boost performance.
• Applied the RandAugment data‑augmentation method to improve model robustness and generalisation.

Data Augmentation Method

• RandAugment: Randomly selects a set of augmentation operations—including rotation, flipping, cropping, scaling, etc.—and randomly chooses intensity parameters, applying them to training samples to increase data diversity.

Experimental Results

• The Attention‑SwinTransformer achieved an accuracy of 96.7% on the test set.
• The SwinTransformer_ResNet_ATT reached 95.73% accuracy on the test set, outperforming other models.
• Data augmentation improved performance across several models, e.g., an 8.11% gain for EfficientNet and a 12.53% gain for SwinTransformer_STN.

Conclusion

• By innovatively improving the SwinTransformer architecture and incorporating an STN module, as well as constructing a multi‑model attention mechanism, this work significantly enhanced performance on leaf‑disease classification in plant pathology.
• The RandAugment augmentation effectively increased model adaptability to input data and improved generalisation.