A dual adaptive semi-supervised attentional residual network framework for urban sound classification

Urban sound classification is essential for efficiently mitigating noise pollution, improving public health, optimizing smart city planning, upgrading mobility, and emergency response systems. Deep learning techniques have demonstrated encouraging outcomes in several sound categorization assignments. However, their implementation in urban data poses difficulties due to the distinctive attributes of urban data, including excessive noise, restricted resolution, and intricate scattering patterns. This study introduces a new ResNet-Attention framework that is specifically tailored for the classification of urban sounds. The system combines the benefits of Residual Networks (ResNet) and attention processes to improve feature extraction and discriminative capability. The ResNet component facilitates the acquisition of profound representations, while the attention mechanism discriminately concentrates on significant regions in the urban data. We assess the proposed framework using benchmark urban datasets, namely Detection and Classification of Acoustic Scenes and Events (DCASE) and compare its performance with the most advanced methods available. The network DASS-ARN1 achieves an outstanding accuracy of 71.18 % on the test dataset by using only 25 % of the available labeled data. This outstanding accuracy represents a considerable improvement of 11.60 % compared to the accuracy reported in the baseline technique. In addition, our alternative network architecture, DASS-ARN2, outperforms these results by reaching a greater accuracy of 72.93 %, representing a significant improvement of 13.35 %. In addition, we perform thorough ablation studies to examine the specific impacts of the ResNet and attention components. The suggested architecture demonstrates the considerable potential for precise and dependable urban sound classification.