Abstract
In the context of birdsong recognition, conventional modeling approaches often involve a significant number of parameters and high computational costs, rendering them unsuitable for deployment in embedded field monitoring devices. To improve the convenience of birdsong recognition, this study proposes a more lightweight model based on the original EfficientNet-B0 architecture. The proposed method introduces the ECA attention mechanism to reduce the parameter complexity while improving feature expression. Furthermore, by adjusting the convolution kernel in the MBConv structure, and incorporating the CBAM attention mechanism in several intermediate layers, we achieve further enhancement in model accuracy. Finally, we employ the Adam optimization algorithm to improve network convergence speed. Our approach attains an impressive 96.04% accuracy in recognizing ten bird species-an improvement of 3.2% over the original model-while reducing parameter count by 16.4%, thereby enhancing both accuracy and convenience.