Sound Event Detection Using EfficientNet-B2 with an Attentional Pyramid Network

This paper proposes a sound event detection (SED) model that uses EfficientNet-B2 and an attentional pyramid network (APNet) module to effectively represent information from a multi-resolution feature map. Compared to the A2FPN-based SED model, the proposed SED model is realized with a reduced computational complexity and improved performance due to the newly proposed APNet. The proposed SED model is based on a pre-trained EfficientNet-B2 that is obtained from the pretraining, sampling, labeling, and aggregation framework. Then, multi-resolution feature maps extracted from EfficientNet-B2 are aggregated by the APNet module. The aggregated feature map is then used to detect sound events by using a detection network mainly composed of two bidirectional gated recurrent unit layers. The performance of the proposed SED model is evaluated on the detection and classification of acoustic scenes and events (DCASE) 2022 Challenge Task 4. Consequently, it was shown that the F1-score and polyphonic sound event detection scores 1 and 2 of the proposed SED model are higher by 0.4%, 0.009, and 0.014, respectively, than those of the A2FPN-based SED model. In addition, the proposed model had a smaller number of floating point operations than the A2FPN-based SED model.