Wild Bird Species Identification Based on a Lightweight Model

Abstract

This work solves the difficulty of accurately
identifying bird species with technologies of sound
identification in environmental and protection.
Conventional models of the convolutional neural
network (CNN) encounter difficulty in decrypting
complex correlations present in the spectrograms,
which prevents their use in real environment due to
high processing requirements. To solve this, we
introduce a light model using frequency dynamic
convolution and maintain the gentle characteristics of
bird vocalization across several frequency bands. The
use of coordinate attention increases the acquisition of
global information and therefore increases the
efficiency of the model. Using several deep learning
architectures, such as Resnet50 and lightweight
Mobilenet, we have achieved commendable results,
especially 96% accuracy with MobileNetV3 Large.
The use of this success, the methodology of the
ensemble has increased accuracy. Our model, which
integrates Mobilenetv3 large with random forest, has
achieved a perfect 100% accuracy, showing the
efficiency of merging deep learning with traditional
machine learning methodologies. This study
illustrates the effectiveness of our compact model for
identifying bird species and provides a scalable
solution for field application and increases studies in
population ecology and protection biology