Tomato leaf disease recognition based on multi-task distillation learning.
Bo Liu1,2, Shusen Wei1,2, Fan Zhang1,2
1College of Information Science and Technology, Hebei Agricultural University, Baoding, China.
Frontiers in Plant Science
|February 14, 2024
Summary
This study introduces a novel multi-task distillation learning (MTDL) framework for accurate tomato leaf disease diagnosis. The MTDL approach enhances classification and severity prediction while significantly reducing model complexity for intelligent agriculture.
Area of Science:
- Agricultural Science
- Computer Vision
- Machine Learning
Background:
- Tomato leaf diseases significantly impact crop yield and quality.
- Automated disease recognition using computer vision faces challenges like varied symptoms, limited data, and complex models.
Purpose of the Study:
- To develop a novel multi-task distillation learning (MTDL) framework for comprehensive tomato leaf disease diagnosis.
- To effectively leverage shared and unique knowledge between disease classification and severity prediction tasks.
Main Methods:
- Implemented a multi-task distillation learning (MTDL) framework.
- Employed knowledge disentanglement, mutual learning, and knowledge integration in a multi-stage strategy.
- Utilized an MTDL-optimized EfficientNet model.
Main Results:
- The MTDL framework improved performance in both classification accuracy and severity estimation compared to single-task models.
- The MTDL-optimized EfficientNet demonstrated superior performance with significantly reduced model complexity (9.46% of parameters compared to ResNet101).
- Achieved 0.68% higher classification accuracy and 1.52% higher severity estimation accuracy.
Conclusions:
- The proposed MTDL framework offers a practical and efficient solution for intelligent agriculture.
- This approach effectively addresses the challenges in automated tomato leaf disease recognition.
