A Network Architecture Based on Feature Fusion for Acoustic Scene Classification

A network architecture based on feature fusion is proposed to enhance representation learning for improving the accuracy of acoustic scene classification (ASC). In the proposed network model, by utilizing feature fusion blocks, global information is successively transferred from the Swin Transformer branch into the convolutional neural network (CNN) branch, while local features from the CNN branch are progressively sent into the patch embedding for enriching local details of the Swin Transformer branch. This bidirectional interaction way makes the CNN and Swin Transformer complement each other, and their local features and global representations are fully fused, which helps improve the accuracy of ASC. The experimental results prove that the proposed method reaches the superior performances over the reference deep learning methods in ASC.