Scholarlink

<p>Aimed at the problems of high complexity and low classification accuracy in the existing methods of acoustic scene classification, an acoustic scene classification method based on time-frequency multi-aggregation and cross Gaussian attention (TFMA-CGA) is proposed. Initially, based on the principle of time-frequency separation convolutions, an algorithm for time-frequency multi-aggregation convolutions (TFMA-Convs) is proposed. The algorithm uses asymmetric convolution structure to achieve time-frequency separation and aggregation, generating a feature map of a first-order time-frequency aggregation. Subsequently, time-frequency dual channel attention is utilized to enhance the key time-frequency information, and then aggregated in the channel dimension to generate a feature map of second-order time-frequency aggregation, thereby enhancing the model’s receptive field and representational capability. Furthermore, the TFMA-CGA is constructed on TFMA-Convs. The network uses feature pyramids and cross Gaussian attention, aiming to achieve hierarchical fusion from shallow to deep while establishing global correlation, thereby improving the model’s classification accuracy. Finally, the acoustic scene classification experiment conducted on UrbanSound8K, DCASE2019, and DCASE2020 datasets. The experimental results showed that the proposed TFMA-CGA method can effectively improve the model’s receptive field, integrate global context information, mitigate feature collapse, and improve the model’s representation ability. The optimal classification accuracy of the model reached 93.78%, 73.76%, and 69.91% on the three datasets respectively, with its model parameters and MFLOPS outperforming many mainstream methods, demonstrating the effectiveness of the proposed approach.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Time-Frequency Multi-Aggregation and Cross Gaussian Attention for Environmental Acoustic Scene Classification

  • 期刊论文

  • 发表日期：

    2026年3月12日

查看全文
  • Yanwen Wang,
  • Yi Cao,
  • Yangyang Shi

摘要

Aimed at the problems of high complexity and low classification accuracy in the existing methods of acoustic scene classification, an acoustic scene classification method based on time-frequency multi-aggregation and cross Gaussian attention (TFMA-CGA) is proposed. Initially, based on the principle of time-frequency separation convolutions, an algorithm for time-frequency multi-aggregation convolutions (TFMA-Convs) is proposed. The algorithm uses asymmetric convolution structure to achieve time-frequency separation and aggregation, generating a feature map of a first-order time-frequency aggregation. Subsequently, time-frequency dual channel attention is utilized to enhance the key time-frequency information, and then aggregated in the channel dimension to generate a feature map of second-order time-frequency aggregation, thereby enhancing the model’s receptive field and representational capability. Furthermore, the TFMA-CGA is constructed on TFMA-Convs. The network uses feature pyramids and cross Gaussian attention, aiming to achieve hierarchical fusion from shallow to deep while establishing global correlation, thereby improving the model’s classification accuracy. Finally, the acoustic scene classification experiment conducted on UrbanSound8K, DCASE2019, and DCASE2020 datasets. The experimental results showed that the proposed TFMA-CGA method can effectively improve the model’s receptive field, integrate global context information, mitigate feature collapse, and improve the model’s representation ability. The optimal classification accuracy of the model reached 93.78%, 73.76%, and 69.91% on the three datasets respectively, with its model parameters and MFLOPS outperforming many mainstream methods, demonstrating the effectiveness of the proposed approach.