IRMKD: an application of instance relation matrix in plant disease recognition

Area of Science:

• Agricultural Science
• Computer Science
• Artificial Intelligence

Background:

• Plant disease recognition is crucial for crop yield.
• Convolutional Neural Networks (CNNs) excel at plant disease identification but are computationally intensive.
• High resource requirements of CNNs limit their deployment on edge devices.

Purpose of the Study:

• To develop an efficient plant disease recognition method for devices with limited computing resources.
• To reduce the model parameters and recognition time of existing plant disease identification systems.
• To enable accurate plant disease detection on low-power devices.

Main Methods:

• Introduced instance-relation-matrix based knowledge distillation (IRMKD).
• Transferred mutual relations between data examples using instance correlations and inter-instance characteristic relationships.
• Employed multiple loss functions for effective knowledge distillation.

Main Results:

• The proposed IRMKD method significantly reduced memory usage and recognition time by an average of 92% for traditional neural networks.
• Recognition accuracy was maintained above 93%.
• Demonstrated improved performance for student models trained with IRMKD.

Conclusions:

• IRMKD effectively reduces model volume and enhances recognition speed.
• The method provides a viable solution for plant disease recognition on devices with limited memory and computational power.
• Offers a new approach for efficient plant disease identification in agriculture.