Skip to content
Conference

Dual-domain synergistic semisupervised sound event classification with semantic protection

Sep 2026 · International Conference on Signal Processing and Communication Security · Vol 14374, pp. 1437405 - 1437405-7 · 0 citations · 17 references
Engineering

Abstract

Semi-supervised learning (SSL) has shown significant potential in reducing the reliance of sound event classification (SEC) on large-scale, high-quality annotated datasets. However, transferring visual SSL frameworks to audio tasks often encounters two primary limitations: first, aggressive data augmentation tends to disrupt continuous time-frequency structures, leading to semantic fragmentation at the feature level; second, under complex noise interference, the exclusive reliance on prediction-layer constraints and fixed thresholds is prone to inducing confirmation bias. To address these challenges, this paper proposes a dual-domain synergistic semi-supervised audio classification architecture. In the prediction domain, a lightweight semantic protection mechanism based on Standard Class Activation Mapping (Standard CAM) is introduced to prevent the loss of core features during augmentation with minimal computational overhead. Concurrently, a global-aware category dynamic threshold strategy is integrated to effectively mine high-confidence hard samples. In the feature domain, a large-capacity queue is constructed to execute instance-level contrastive learning, which aligns homologous positive samples and maximizes their distance from historical negative samples in the latent representation space, thereby further enhancing the model's underlying noise-robust representation capability. Experimental results on the public ESC-50 and UrbanSound8K datasets demonstrate that the proposed method maintains stable classification performance across various label ratios. Compared with mainstream baselines, the proposed model achieves accuracy improvements of up to 2.5% and 6.8% on the UrbanSound8K and ESC-50 datasets, respectively, verifying the effectiveness and superiority of the dual-domain synergistic architecture in complex acoustic environments.

View source

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.