Improved FasterViT model for citrus disease diagnosis

Area of Science:

• Computer Vision
• Plant Pathology
• Machine Learning

Background:

• Deep learning models struggle with plant disease recognition due to environmental variations and challenges in merging local/global image information, especially with small datasets.
• Traditional methods often lead to impaired performance and high training costs in plant disease identification tasks.

Purpose of the Study:

• To propose an improved FasterViT model for accurate citrus disease recognition, focusing on efficient learning from small-scale datasets.
• To enhance model robustness, generalization, and reduce training costs in plant disease detection.

Main Methods:

• Developed an improved FasterViT model, a hybrid Convolutional Neural Network (CNN)-Vision Transformer (ViT) framework, integrating CNN's local feature extraction with ViT's global context understanding.
• Implemented cross-stage alternating Mixup and Cutout for data augmentation, alongside Triplet Attention and AdaptiveAvgPool mechanisms to optimize performance and reduce computational load.
• Validated the model on a custom in-field small citrus disease dataset and the PlantVillage dataset.

Main Results:

• The improved FasterViT model demonstrated fast learning capabilities and effective adaptation to small sample training for plant disease detection.
• Achieved significant improvements in model accuracy and notable reductions in training costs compared to baseline models.
• Showcased excellent performance in transfer learning scenarios, highlighting adaptability and broad applicability.

Conclusions:

• The proposed improved FasterViT model effectively addresses the complexities of plant disease image recognition, particularly for small-scale datasets.
• This approach offers an efficient, scalable, and robust classification system for plant disease detection.
• The study pioneers a new paradigm for developing advanced deep learning solutions in agricultural applications.