Class Activation Mapping-Driven Data Augmentation: Masking Significant Regions for Enhanced Acoustic Scene Classification

This paper proposes a novel masking augmentation method for acoustic scene data using class activation mapping (CAM) to improve the performance of audio scene classification (ASC) tasks. CAM enables the identification of significant regions within the input data by emphasizing areas exhibiting the strongest activation, obtained through gradients resulting from backpropagation with respect to specific classes. By masking these important regions identified by CAM, discriminative features can be learned from less important areas, thereby mitigating the overfitting to the important regions. The proposed augmentation method promotes deep learning model robustness and generalization, and mitigates potential performance fluctuations caused by random masking. Furthermore, we investigate various masking shapes, such as rectangular, cross-shaped, and circular masks. Experimental results on the DCASE challenge 2019 Dataset demonstrate that applying masking techniques based on CAM algorithms leads to superior ASC accuracy compared to SpecAugment, which is the most widely used masking augmentation method in combination with two ResNet-based model architectures. This study highlights the potential of CAM-driven data augmentation in the ASC domain and provides valuable insights for future research and development in the field.