# Enhancing Sound Texture in CNN-Based Acoustic Scene Classification

**Authors:** Yuzhong Wu, Tan Lee  
**arXiv:** [1901.01502](https://arxiv.org/abs/1901.01502)  
**Date:** 2019-01-06  
**Summary page:** https://www.summarizepaper.com/en/paper/1901.01502v1/

## Key points

- Acoustic scene classification involves identifying the scene from which an audio signal is recorded
- Convolutional neural network (CNN) models have been successful for this task
- Limited understanding of how audio scenes are perceived in CNN compared to image recognition research
- Class Activation Mapping (CAM) used to analyze how log-Mel features of different acoustic scenes are learned in a CNN classifier
- High-energy time-frequency components do not necessarily correspond to strong activation on CAM, instead background sound texture is well learned
- Difference of Gaussian (DoG) and Sobel operator applied to process log-Mel features, enhancing edge information in time-frequency image representation
- Evaluation on DCASE 2017 ASC challenge dataset shows that using edge-enhanced log-Mel images as input features significantly improves classification performance
- Study provides insights into how acoustic scenes are perceived in CNNs and proposes a method to enhance sound texture representation for improved accuracy
- Findings contribute to advancing acoustic scene classification techniques with potential applications in surveillance systems or environmental monitoring.

## Summary

The paper titled "Enhancing Sound Texture in CNN-Based Acoustic Scene Classification" by Yuzhong Wu and Tan Lee explores the task of acoustic scene classification, which involves identifying the scene from which an audio signal is recorded. Convolutional neural network (CNN) models have been widely adopted for this task and have shown success. However, there is limited understanding of how an audio scene is perceived in a CNN compared to image recognition research. To address this gap, the authors utilize Class Activation Mapping (CAM) to analyze how log-magnitude Mel-scale filter-bank (log-Mel) features of different acoustic scenes are learned in a CNN classifier. They find that distinct high-energy time-frequency components of audio signals do not necessarily correspond to strong activation on CAM. Instead, the background sound texture is well learned by the CNN. In order to make the sound texture more salient and improve classification performance, the authors propose applying the Difference of Gaussian (DoG) and Sobel operator to process the log-Mel features. This enhances edge information in the time-frequency image representation. The authors evaluate their approach on the DCASE 2017 ASC challenge dataset and demonstrate that using edge-enhanced log-Mel images as input features for CNN significantly improves audio scene classification performance. Overall, this study provides valuable insights into how acoustic scenes are perceived in CNNs and proposes an effective method to enhance sound texture representation for improved classification accuracy. The findings contribute significantly to advancing acoustic scene classification techniques and can potentially benefit various applications such as surveillance systems or environmental monitoring.