Evaluation of Deep Neural Network Compression Methods for Edge Devices Using Weighted Score-Based Ranking Scheme.

Area of Science:

• Computer Vision
• Edge Computing
• Machine Learning Model Optimization

Background:

• The increasing demand for object detection in edge computing necessitates lightweight Convolutional Neural Network (CNN) models.
• Current CNN models often exceed memory constraints for edge device deployment, requiring optimization without performance loss.

Purpose of the Study:

• To rank the efficiency of various model compression methods for edge devices.
• To develop a real-time vehicle tracking system for cargo ships as a case study.
• To propose a weighted score-based ranking scheme for selecting optimal compression techniques.

Main Methods:

• Developed a weighted score-based ranking scheme using model performance metrics.
• Applied the ranking scheme to baseline, compressed, and micro-CNN models.
• Evaluated models on a dataset for a real-time vehicle tracking application on cargo ships.

Main Results:

• Quantization emerged as the most efficient compression method, achieving the highest rank with an average weighted score of 9.00.
• Binarization followed as the second most effective method, with an average weighted score of 8.07.
• The proposed ranking scheme effectively identified the optimal compression strategy for the specific application.

Conclusions:

• Quantization is the most suitable compression technique for developing lightweight CNNs for real-time vehicle tracking on edge devices.
• The developed weighted score-based ranking scheme provides a flexible framework for selecting model compression methods across various edge computing applications.
• This research addresses the critical need for efficient model optimization in resource-constrained edge environments.