Multi-Layer CAM and Deep-Shallow Feature Fusion CNN for Weakly Supervised Sound Event Detection

The data required for training a supervised Sound Event Detection (SED) model needs to be manually annotated, which makes acquiring it quite difficult. During the process of obtaining temporal localization from weakly supervised networks using Class Activation Map (CAM), due to the low resolution of CAM, the accuracy of the positioning is poor. In this paper, we propose a weakly supervised SED method based on a deep-shallow feature fusion CNN. It extracts feature maps from multiple convolutional layers to capture the network's focus areas on features, and constructs a contrastive loss for training to make the network focus on features more accurately. Furthermore, aggregating the CAMs extracted from different convolutional layers with weighted aggregation further improves the performance of the method proposed in this paper. Experiments on the Domestic environment sound event detection (DESED) dataset demonstrate that the proposed method achieves promising performance.