AgriKD: Cross-Architecture Knowledge Distillation for Efficient Leaf Disease Classification
Researchers have introduced AgriKD, a novel cross-architecture knowledge distillation framework designed to enable efficient leaf disease classification on resource-constrained edge devices. While Vision Transformers (ViTs) offer superior representation capabilities for detecting plant diseases, their high computational costs hinder deployment in agricultural field environments. AgriKD addresses this by transferring knowledge from a heavy ViT teacher model to a lightweight convolutional student model. The framework bridges the architectural gap by integrating distillation objectives at output, feature, and relational levels, allowing the student model to effectively capture global representations. Experimental results demonstrate that the distilled student model achieves performance comparable to the teacher while reducing model parameters by approximately 172 times and computational cost by 47.57 times. Inference latency is improved by 18-22 times. The optimized model supports multiple runtime formats, including ONNX and TensorRT, and has been successfully deployed on NVIDIA Jetson devices and mobile applications, proving its practicality for real-time, AI-powered agricultural monitoring.
Wire timeline
AgriKD: Cross-Architecture Knowledge Distillation for Efficient Leaf Disease Classification
Researchers have introduced AgriKD, a novel cross-architecture knowledge distillation framework designed to enable efficient leaf disease classification on resource-constrained edge devices. While Vision Transformers (ViTs) offer superior representation capabilities for detecting plant diseases, their high computational costs hinder deployment in agricultural field environments. AgriKD addresses this by transferring knowledge from a heavy ViT teacher model to a lightweight convolutional student model. The framework bridges the architectural gap by integrating distillation objectives at output, feature, and relational levels, allowing the student model to effectively capture global representations. Experimental results demonstrate that the distilled student model achieves performance comparable to the teacher while reducing model parameters by approximately 172 times and computational cost by 47.57 times. Inference latency is improved by 18-22 times. The optimized model supports multiple runtime formats, including ONNX and TensorRT, and has been successfully deployed on NVIDIA Jetson devices and mobile applications, proving its practicality for real-time, AI-powered agricultural monitoring.
cs.AI updates on arXiv.org