计算机科学
卷积神经网络
预处理器
人工智能
噪音(视频)
推论
突出
计算机视觉
模式识别(心理学)
干扰(通信)
卷积(计算机科学)
GSM演进的增强数据速率
环境噪声
深度学习
机器学习
特征提取
语音识别
数据预处理
数据挖掘
边缘设备
人工神经网络
信号(编程语言)
自编码
信号处理
光谱图
噪声测量
作者
Pengshun Che,Jiaan Liu,Lu Jiang
出处
期刊:
日期:2025-11-24
卷期号:: 468-472
标识
DOI:10.1109/isceic67854.2025.11405525
摘要
Urban sound classification plays a crucial role in smart city applications, enabling noise monitoring, emergency detection, and environmental analysis. This paper proposes a lightweight Convolutional Neural Network (CNN) model tailored for classifying urban sounds using Mel-spectrograms derived from a subset of the UrbanSound8K dataset. The model emphasizes efficient signal preprocessing via Short-Time Fourier Transform (STFT) and Mel-scale filtering to capture salient frequency-time features, achieving a test accuracy of 86.53% on a 1000-sample subset while utilizing only 181k parameters. Inspired by self-supervised multisensory features from ICCV 2019, our approach integrates data augmentation and shallow convolutional layers to balance performance and computational efficiency. Extensive experiments demonstrate superior results over baselines like ResNet-18 and MFCC+SVM, with ablation studies validating the contributions of preprocessing and augmentation. The lightweight design ensures rapid training (12.13 minutes) and inference (0.237 seconds), making it suitable for edge devices. This work advances accessible urban acoustics analysis, paving the way for real-time deployments in resource-constrained environments.
科研通智能强力驱动
Strongly Powered by AbleSci AI