Page 60 - 《应用声学》2025年第3期

594
2.5 本文方法与其他方法的性能对比
表 5 是本文方法与一些经典方法的对比，包
括基于 VGG [23] 、CRNN [24] 、ResNet [16] 和 Trans-
former [25] 典型神经网络结构的分类方法。相较
于这些经典模型，DADHPN 在分类准确率上表现
出色，达到了89.5%，而应用 EIL 后的分类准确率为
90.1%。DADHPN 结合了注意力机制、双层级并行
分类以及教师 -学生模型的知识传递策略，充分利
用了高层级场景特征知识的辅助，动态调整低层级
场景通道特征权重，提高了低层级场景的分类性能。
同时，EIL 结合了高层级场景预测和低层级场景预
测结果，增强了网络的低层级场景辨识能力，进一步
提升了分类准确率。
表 5 模型对比分类指标
Table 5 Model comparison classification
indicators
模型
VGG-like CNN [23]
CNN-GRU [24]
ResNet [16]
Transformer Encoder [25]
DADHPN
DADHPN + EIL
3 结论
本文提出了一种基于注意力的双层级并行网
络(DADHPN)用于 ASC，其综合了注意力机制、双
层级并行分类和教师 -学生模型的知识传递策略。
DADHPN 能够动态调整学生模型的学习焦点，利
用教师模型知识实现对学生模型特征的选择性关
注，提高了学生模型的场景分类能力，使其分类准
确率达到了89.5%。此外，通过EIL融合高低层级分
类模型的输出，弥补了学生模型单一预测的局限性，
提高了学生模型在低层级场景分类方面的能力。最
终，获得了 90.1% 的分类准确率，超越了一些典型
ASC网络的性能，验证了本文方法的有效性。
参 考 文 献
[1] 李伟, 李硕. 理解数字声音 ——基于一般音频/环境声的
计算机听觉综述 [J]. 复旦学报 (自然科学版), 2019, 58(3):
269–313.
2025 年 5 月
Li Wei, Li Shuo. Understanding digital audio——A re-
view of general audio/ambient sound based computer au-
dition[J]. Journal of Fudan University (Natural Science),
2019, 58(3): 269–313.
[2] 陈爱武. 家居音频场景识别关键技术研究 [D]. 广州: 华南理
工大学, 2020: 2.
[3] Chu S, Narayanan S, Kuo C C J. Environmental sound
recognition with time–frequency audio features[J]. IEEE
Transactions on Audio, Speech, and Language Processing,
2009, 17(6): 1142–1158.
[4] Phaye S S R, Benetos E, Wang Y. Subspectralnet–using
sub-spectrogram based convolutional neural networks for
acoustic scene classification[C]// 2019 IEEE International
Conference on Acoustics, Speech and Signal Processing
(ICASSP). IEEE, 2019: 825–829.
[5] McDonnell M D, Gao W. Acoustic scene classification us-
ing deep residual networks with late fusion of separated
high and low frequency paths[C]// 2020 IEEE Interna-
tional Conference on Acoustics, Speech and Signal Pro-
cessing (ICASSP). IEEE, 2020: 141–145.
[6] Chen H, Liu Z, Liu Z, et al. Integrating the data augmen-
tation scheme with various classifiers for acoustic scene
准确率/%
modeling[J]. arXiv Preprint, arXiv: 1907.06639, 2019.
81.6 [7] Singh A, Rajan P, Bhavsar A. Deep multi-view features
88.1 from raw audio for acoustic scene classification[C]// IEEE
88.4 AASP Challenge on Detection and Classification of Acous-
tic Scenes and Events (DCASE), 2019: 229–233.
88.9
[8] Kim T, Lee J, Nam J. Comparison and analysis of Sam-
89.5 pleCNN architectures for audio classification[J]. IEEE
90.1 Journal of Selected Topics in Signal Processing, 2019,
13(2): 285–297.
[9] Kong Q, Cao Y, Iqbal T, et al. Panns: Large-scale pre-
trained audio neural networks for audio pattern recogni-
tion[J]. IEEE/ACM Transactions on Audio, Speech, and
Language Processing, 2020, 28: 2880–2894.
[10] Nigro M, Rueda A, Krishnan S. Acoustic scene classifi-
cation using time–frequency energy emphasis and convo-
lutional recurrent neural networks[C]// Artificial Intelli-
gence and Evolutionary Computations in Engineering Sys-
tems: Computational Algorithm for AI Technology, Pro-
ceedings of ICAIECES 2020. Springer Singapore, 2022:
267–276.
[11] Hou Y, Kang B, Van Hauwermeiren W, et al. Relation-
guided acoustic scene classification aided with event em-
beddings[C]// 2022 International Joint Conference on
Neural Networks (IJCNN). IEEE, 2022: 1–8.
[12] Huang J, Lu H, Lopez Meyer P, et al. Acoustic scene
classification using deep learning-based ensemble averag-
ing[C]// IEEE AASP Challenge on Detection and Clas-
sification of Acoustic Scenes and Events (DCASE), 2019:
94–98.
[13] Ren Z, Kong Q, Han J, et al. Attention-based atrous
convolutional neural networks: Visualisation and under-
standing perspectives of acoustic scenes[C]// 2019 IEEE
International Conference on Acoustics, Speech and Signal