Implementing a Compression Technique on the Progressive Contextual Excitation Network for Smart Farming Applications

Area of Science:

• Computer Vision
• Deep Learning
• Agricultural Automation

Background:

• Smart farming increasingly utilizes computer vision for automation.
• Deep learning models excel in image classification and object detection.
• The Progressive Contextual Excitation Network (PCENet) shows promise for agricultural image analysis.

Purpose of the Study:

• To accelerate the PCENet model for edge computing in smart farming.
• To investigate the effectiveness of pruning filters for model compression.
• To evaluate the trade-off between speed and accuracy.

Main Methods:

• Applied filter pruning technique to compress the PCENet model.
• Evaluated the compressed PCENet model on cocoa bean image classification.
• Assessed performance on the Jetson Nano edge platform.
• Tested the compressed model on a corn leaf disease dataset.

Main Results:

• Achieved a processing speed of 16.7 FPS on Jetson Nano, an improvement from 9.9 FPS.
• Maintained high accuracy for cocoa bean classification (86.1% compressed vs. 86.8% original).
• Outperformed ResNet18 (82.7%) in accuracy for cocoa bean classification.
• Achieved 97.5% accuracy on the corn leaf disease dataset post-compression.

Conclusions:

• Filter pruning effectively accelerates PCENet for real-time smart farming applications.
• The compressed PCENet model offers a favorable balance between computational efficiency and classification accuracy.
• The approach demonstrates potential for deploying advanced deep learning models on resource-constrained edge devices.