Enhancing Environmental Sound Classification with Multi-Resolution Feature Fusion and Hybrid Attention Networks

To address the limitations of single-frequency resolution in time-frequency features for accurately capturing the characteristics of diverse sound types, this paper presents a novel deep learning-based approach for sound classification. This method enhances data representation and captures more contextual information during the feature extraction process. It synthesizes multi-frequency resolution features with a ResNet-based hybrid attention network, thereby realizing innovation in both time-frequency feature extraction and classification model design. In this paper, multi-frequency resolution features (MRF) based on three audio feature extraction techniques are introduced for feature representation, which improves the model's ability to capture the diversity of sound signals through the combination of different frame lengths. Meanwhile, relying on the powerful learning capability of deep residual networks (ResNet), this paper introduces a hybrid attention mechanism containing channel and feature attention modules to effectively filter out background noise and redundant information. In this paper, the UrbanSound8K dataset is used to compare the proposed model with the hybrid model based on deep learning. The experimental results show that the ESC model proposed in this paper achieves a high classification accuracy of 98.4 % on the UrbanSound8K dataset, which outperforms existing classification methods.