An improved lightweight method based on EfficientNet for birdsong recognition.

Area of Science:

• Bioacoustics
• Machine Learning
• Computer Vision

Background:

• Conventional birdsong recognition models are computationally intensive and parameter-heavy, limiting their use in embedded field devices.
• There is a need for efficient and accurate birdsong recognition systems for ecological monitoring.

Purpose of the Study:

• To develop a lightweight and accurate birdsong recognition model suitable for embedded field monitoring.
• To improve the efficiency and feature expression of birdsong recognition models.

Main Methods:

• A modified EfficientNet-B0 architecture was developed, incorporating the ECA attention mechanism to reduce parameters and enhance feature expression.
• The MBConv structure's convolution kernel was adjusted, and the CBAM attention mechanism was integrated into intermediate layers.
• The Adam optimization algorithm was utilized to accelerate network convergence.

Main Results:

• The proposed model achieved 96.04% accuracy in recognizing ten bird species, a 3.2% improvement over the original model.
• Parameter count was reduced by 16.4%, making the model more suitable for embedded systems.
• Enhanced feature expression and faster convergence were observed.

Conclusions:

• The lightweight EfficientNet-B0 model offers a significant improvement in birdsong recognition accuracy and efficiency.
• The model's reduced complexity and enhanced performance make it ideal for practical applications in ecological field monitoring.
• This approach successfully balances accuracy and computational cost for birdsong recognition.