MFF-Net: A multi-scale feature fusion network for birdsong classification

Abstract

In this paper, we propose a novel birdsong classification network, MFF-Net(Multi-scale Feature Fusion Network), which enhances classification performance through multi-scale feature fusion. The network is composed of four components. The first one is a multi-scale feature extraction module that extracts different scale features from the original sound. The second one is a feature fusion module utilizing a channel attention mechanism to integrate these features effectively. The third one is a feature replacement module designed to replace low-weight features and enhance feature representation. And the fourth one is a classifier module that performs birdsong classification. The proposed method was evaluated on two publicly available birdsong datasets and an urban sound dataset(Urbansound8k) to test its generalization performance. Experimental results showed that MFF-Net achieved a classification accuracy of 96.83 % on the BirdCLEF-13 dataset and demonstrated good generalization performance on the urban sound dataset (UrbanSound8k), achieving competitive results. These results highlight the robustness and effectiveness of MFF-Net in noisy and diverse environments.