Acoustic Scene Classification Using Pixel-Based Attention

Wang X, Xu Y, Shi J, Teng X. Acoustic Scene Classification Using Pixel-Based Attention. J Audio Eng Soc. 2020;68(11):843-855. doi:10.17743/jaes.2020.0052

@article{Wang2020_20998,
  author = {Wang, Xingmei and Xu, Yichao and Shi, Jiahao and Teng, Xuyang},
  title = {{Acoustic Scene Classification Using Pixel-Based Attention}},
  journal = {Journal of the Audio Engineering Society},
  volume = {68},
  number = {11},
  pages = {843--855},
  year = {2020},
  month = nov,
  publisher = {Audio Engineering Society},
  doi = {10.17743/jaes.2020.0052},
  url = {https://doi.org/10.17743/jaes.2020.0052}
}

TY  - JOUR
TI  - Acoustic Scene Classification Using Pixel-Based Attention
AU  - Wang, Xingmei
AU  - Xu, Yichao
AU  - Shi, Jiahao
AU  - Teng, Xuyang
T2  - Journal of the Audio Engineering Society
J2  - J. Audio Eng. Soc.
VL  - 68
IS  - 11
SP  - 843
EP  - 855
PY  - 2020
DA  - 2020/11/06
DO  - 10.17743/jaes.2020.0052
UR  - https://doi.org/10.17743/jaes.2020.0052
PB  - Audio Engineering Society
LA  - en
AB  - In this paper, we propose a pixel-based attention (PBA) module for acoustic scene classification (ASC). By performing feature compression on the input spectrogram along the spatial dimension, PBA can obtain the global information of the spectrogram. Besides, PBA applies attention weights to each pixel of each channel through two convolutional layers combined with global information. In addition, the spectrogram applied after the attention weights is multiplied by the gamma coefficient and superimposed with the original spectrogram to obtain more effective spectrogram features for training the network model. Furthermore, this paper implements a convolutional neural network (CNN) based on PBA (PB-CNN) and compares its classification performance on task 1 of Detection and Classification of Acoustic Scenes and Events (DCASE) 2016 Challenge with CNN based on time attention (TB-CNN), CNN based on frequency attention (FB-CNN), and pure CNN. The experimental results show that the proposed PB-CNN achieves the highest accuracy of 89.2% among the four CNNs, 1.9% higher than that of TB-CNN (87.3%), 2.2% higher than that of FB-CNN (86.6%), and 3% higher than that of pure CNN (86.2%). Compared with DCASE 2016’s baseline system, the PB-CNN improved by 12%, and its 89.2% accuracy was the highest among all submitted single models.
ER  -