CADP: Connection-Aware DenseNet Pruning for lightweight plant disease classification.

Area of Science:

• Agricultural Science
• Computer Science
• Deep Learning

Background:

• Plant diseases pose a significant threat to global food security.
• Deep learning models like DenseNet are effective for plant disease recognition but are computationally intensive.
• Deployment of large models on resource-constrained edge devices is challenging.

Purpose of the Study:

• To develop an efficient compression method for DenseNet models for plant disease recognition.
• To enable the deployment of accurate plant disease recognition models on edge devices.

Main Methods:

• Proposed Connection-Aware DenseNet Pruning (CADP) with three modules: EdgePrune, connection-guided CP decomposition, and dual-stream knowledge distillation.
• EdgePrune models inter-channel feature flows using dual-channel importance scoring.
• Connection-guided CP decomposition uses Connection Importance Index (CII) for adaptive layer compression.
• Dual-stream knowledge distillation recovers losses through soft labels and spatial attention transfer.

Main Results:

• Achieved 88% parameter reduction and 89% computational savings on DenseNet-121.
• Maintained high accuracy: 99.67% on PlantVillage and 99.66% on RiceLeaf datasets.
• Demonstrated competitive accuracy with significantly reduced model size.

Conclusions:

• CADP offers an effective approach for compressing DenseNet models for plant disease recognition.
• The method facilitates practical deployment of accurate AI models on resource-constrained edge devices.
• CADP shows potential for generalizability and practical value in agricultural applications.