OPTIMIZING WILD BIRD SPECIES IDENTIFICATION: INTEGRATING A LIGHTWEIGHT MODEL WITH A HYBRID RANDOM FO

Abstract

Abstract: The sound recognition technology is used to count bird species population to provide accurate data for population ecology and conservation biology research. Bird species can be identified by spectrogram of bird’s sounds. The existing Convolutional Neural Network (CNN) is not good at mining the relationship between features in time-frequency tasks, and the existing CNN has more parameters and complicated calculation, so existing models are not suitable for deployment in an actual field environment. In order to fill this gap, this paper proposes a lightweight model with frequency-dynamic convolution for bird species identification. We use frequency dynamic convolution innovatively to better capture features of bird sounds at different frequencies. First of all, we replaced two-dimensional convolution with frequency­dynamic convolution in order to achieve not-shift invariance of the bird sound's spectrogram, so that we can effectively capture the feature differences of spectrogram in different frequency bands. Then, we replaced part of the Squeeze and Excitation (SE) attention mechanism with the Coordinate Attention (CA) attention mechanism in order to get more comprehensive global information. Finally, the feature fusion module was used to fuse the local and global features. In addition, we built a dataset containing 160 bird sounds, improving the generalization ability of the model. The experimental results show that our model has good generalization ability and is superior to the existing Lightweight CNN, and obtained better results in topi accuracy and top5 accuracy.

Specification

Field o f Invention: Deep Learning
This invention presents a lightweight model for bird species identification using sound spectrograms, addressing the limitations of traditional CNNs in time-frequency tasks. The proposed model uses frequency-dynamic convolution to better capture feature differences across various frequency bands, enhancing identification accuracy. A Coordinate Attention (CA) mechanism replaces parts of the Squeeze and Excitation (SE) attention mechanism to -improve global feature extraction without increasing complexity. Additionally, a feature fusion module integrates local and global features, boosting overall performance. Tested on a dataset of 160 bird sounds, the model shows superior generalization and accuracy, making it suitable for real-time field deployment in ecological research and conservation.
Background Art including citations o f prior art: Prior work in bird species identification has utilized CNNs for analyzing bird call spectrograms, as seen in "Bird Song Classification using CNN-based Spectrogram Analysis" (Stowell et al., 2018). However, traditional CNNs struggle with frequency variations and computational complexity, limiting their field deployment. Efforts to create lightweight models, like in "Lightweight Convolutional Neural Networks for Real-Time Acoustic Bird Monitoring" (Grill et al., 2019), aimed to address these issues but faced challenges in complex environments.
Advanced techniques, such as frequency-sensitive convolutions (Choi et al., 2020) and attention mechanisms (Zhu et al., 2021), improved accuracy but often increased computational demands. This invention innovates by integrating frequency-dynamic convolution and a lightweight Coordinate Attention mechanism to enhance feature extraction while maintaining
efficiency.
Objective o f invention (the invention's objectives and advantages, or alternative em bodim ents of the invention): The primary objective of the invention is to develop an efficient and accurate bird species

identification system using audio data, optimized for deployment in field environments with limited computational resources. The specific objectives are: 1. Enhance Feature Extraction: Utilize frequency-dynamic convolution to improve the identification accuracy by capturing fine-grained differences across various frequency bands in bird sound spectrograms. This addresses the limitations of traditional CNNs in
handling time-frequency shifts.
2. Improve Computational Efficiency: Replace parts of the traditional Squeeze and Excitation (SE) attention mechanism with a lightweight Coordinate Attention (CA) mechanism, which captures comprehensive global information without significantly increasing computational complexity, making the model suitable for real-time
applications.
3. Achieve Robust and Generalized Performance: Integrate a feature fusion module to merge local and global features, enhancing the model's ability to distinguish between similar bird species across different environmental conditions. 4. Field Deployment Feasibility: Design a lightweight and scalable model that can be effectively deployed on low-resource devices like mobile phones or edge computing platforms, enabling real-time bird monitoring in diverse ecological settings.
Advantages of the Invention
• Higher Accuracy: Superior bird species identification accuracy with better top-1 and top-5 results compared to existing lightweight CNNs. • Efficient Resource Utilization: Reduced computational demands allow the model to be implemented on low-power devices, suitable for real-time monitoring in the wild. • Versatile Application: Capable of distinguishing a wide range of bird species in varying environmental conditions, enhancing the reliability of data for ecological studies. • Scalable Dataset Utilization: Built on a dataset of 160 bird sound samples, showcasing robust generalization, and the potential to scale with larger datasets without sacrificing

performance.
Alternative Embodiments
1. Modular Attention Mechanisms: Experimenting with other attention mechanisms (e.g., Convolutional Block Attention Module) to further reduce complexity while maintaining
global information accuracy.
2. Integration with Mobile Platforms: Adapting the model for direct integration into mobile applications for citizen science initiatives, making bird species identification accessible
to the general public.
3. Hybrid Ensemble Approach: Combining the lightweight model with other ensemble techniques (e.g., Random Forest or Decision Trees) to enhance robustness in noisy audio environments or when dealing with low-quality field recordings.


Summary of Invention:
The invention presents a novel approach to bird species identification using sound recognition technology, addressing the challenges posed by traditional models in terms of computational complexity and accuracy. It introduces a lightweight model that leverages frequency-dynamic convolution to effectively capture the unique features of bird sounds across various frequency bands. This innovative method enhances the model's ability to identify species with greater precision while maintaining a low computational footprint, making it suitable for real-time applications in resource-constrained environments.
Additionally, the invention incorporates a Coordinate Attention (CA) mechanism to replace portions of the traditional Squeeze and Excitation (SE) attention framework, allowing for improved extraction of global features without significantly increasing processing demands. A feature fusion module further • enhances performance by integrating both local and global features, leading to more robust identification capabilities.
The model is built on a comprehensive dataset containing 160 distinct bird sounds/which not
AirMI
only improves generalization but also ensures effective deployment in diverse ecological
settings. Experimental results demonstrate that this lightweight model outperforms existing CNN-based approaches in both to

p-1 and top-5 accuracy metrics, making it a valuable tool for
researchers and conservationists.
11 -Nov-2024/135132/202441086595/Form 2(Title Page)
Detailed description of the invention:
This invention focuses on developing an advanced system for identifying wild bird species based on audio signals, specifically utilizing their unique sounds. The system employs a lightweight model that integrates innovative techniques to enhance accuracy while maintaining efficiency, making it suitable for deployment in real-world environments with limited
computational resources.

1. Core Components of the Invention A. Frequency-Dynamic Convolution
• Overview: Traditional convolutional layers in CNNs typically operate in a fixed manner, treating all frequency bands equally. This invention replaces standard 2D convolutions with frequency-dynamic convolution.
• Functionality: This novel convolution method dynamically adjusts its parameters based on frequency characteristics, allowing the model to maintain shift invariance. By doing so, it effectively captures and differentiates the unique features of bird sounds in various
frequency bands.
• Advantages: This leads to improved identification of subtle variations in bird calls that are critical for distinguishing between species, especially those that are acoustically
similar.
B. Coordinate Attention Mechanism (CA)
• Ovendew: The invention integrates a lightweight Coordinate Attention (CA) mechanism in place of certain components of the Squeeze and Excitation (SE) attention mechanism.
£
• Functionality: The CA mechanism enhances the model's ability to focus on relevant

features by incorporating both spatial and channel-wise attention, allowing it to capture global contextual information effectively. • Advantages: This results in a more- comprehensive understanding of the sound spectrograms, improving the model's performance without significantly increasing the
computational burden.
C. Feature Fusion Module
• Overview: To further enhance the model's capability, a feature fusion module is employed to combine local and global features extracted from the audio data. • Functionality: This module integrates low-level features (captured by the frequency­dynamic convolution) and high-level contextual information (obtained through the CA
mechanism).
• Advantages: The fusion of these features leads to a more robust representation of the audio data, facilitating accurate classification of bird species.
2. Dataset Development
• Dataset Composition: The model is trained and evaluated on a specially constructed dataset containing 160 distinct bird sounds. This dataset is designed to cover a wide range of species and acoustic variations, ensuring that the model is exposed to diverse
examples during training.
• Generalization Capability: By incorporating a diverse dataset, the invention enhances the model's ability to generalize across different environmental conditions and recording qualities, thereby increasing its robustness in real-world applications.
3. Implementation and Deployment
• Lightweight Architecture: The overall architecture of the model is designed to be lightweight, ensuring that it can be deployed on devices with limited computational power, such as mobile phones or low-cost embedded systems. • • Real-Time Processing: The model's efficiency allows for real-time processing of audio

data, enabling immediate identification of bird species in the field. • User-Friendly Interface: Future embodiments of the invention may include a mobile application that allows users to record bird sounds and receive instant identification results, fostering citizen science and public engagement in ecological monitoring.
4. Experimental Validation
• Performance Metrics: The proposed model has been evaluated using standard metrics, including Top-1 and Top-5 accuracy. Experimental results demonstrate that the model significantly outperforms existing lightweight CNN approaches, highlighting its effectiveness in identifying bird species accurately. • Comparative Analysis: The invention includes a comparative analysis with other state- of-the-art methods, showcasing improvements in accuracy, generalization ability, and processing speed, thereby validating its advantages.

Claims:
-Nov-2024/135132/202441086595/Form 2(Title Page)
1. A method for identifying wild bird species using audio data through frequency-dynamic convolution and a lightweight architecture, incorporating a Coordinate Attention mechanism for improved feature extraction.
2. The method of claim 1, where frequency-dynamic convolution adjusts parameters based on audio frequency characteristics for enhanced shift invariance. 3. The method of claim 1, wherein the Coordinate Attention mechanism provides spatial and channel-wise focus, optimizing relevant feature extraction with low computational demands. 4. A system implementing the method of claim 1, including a real-time audio processing module and a classification engine for field identification of bird species.

5. The system of claim 4, designed for mobile devices, enabling users to identify bird species in various ecological environments.
6. The method of claim 1, further including a feature fusion module that integrates local and global features for improved classification robustness.
7. The system of claim 5, utilizing a dataset of 160 distinct bird sound samples to enhance generalization and accuracy in species identification.
8. The method of claim 1, evaluating model performance with top-1 and top-5 accuracy metrics, showing superior capabilities over traditional lightweight CNN approaches.

Documents