A Comparative Study of Preprocessing and Model Compression Techniques in Deep Learning for Forest Sound...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Signal Processing

Background:

• Deep learning models excel in complex tasks but often require substantial computational resources.
• Deploying advanced AI on resource-constrained edge devices presents significant challenges.
• Acoustic data analysis benefits from efficient deep learning techniques for real-world applications.

Purpose of the Study:

• To comparatively analyze the performance of seven Convolutional Neural Network (CNN) models for deployment on edge devices.
• To investigate the effectiveness of data augmentation, feature extraction, and model compression techniques.
• To evaluate CNN models using acoustic data from the forest sound dataset.

Main Methods:

• Comparative analysis of seven distinct CNN architectures.
• Application of data augmentation and feature extraction techniques.
• Implementation of model compression strategies including weight/filter pruning and 8-bit quantization.

Main Results:

• Optimized CNNs achieved a balance between accuracy and model size through compression.
• MobileNet-v3-small and ACDNet demonstrated high accuracy (87.95% and 85.64%) with compact sizes (243 KB and 484 KB).
• Weight and filter pruning followed by 8-bit quantization proved effective for model compression.

Conclusions:

• Convolutional Neural Networks can be effectively compressed and optimized for deployment on resource-constrained edge devices.
• The study demonstrates the feasibility of using optimized CNNs for real-time forest environment sound classification.
• Efficient deep learning models are crucial for advancing AI capabilities in edge computing applications.