Dynamic Feature Learning with Involution and Convolution for Predominant Instrument Recognition in Polyphonic Music

The article delves into the challenges of identifying predominant instruments in polyphonic music, where overlapping sounds complicate the isolation and distinction of individual instruments. Traditional methods, such as sliding window analysis and aggregation strategies, are computationally intensive and often require complex post-processing. The article introduces a novel hybrid deep learning framework that integrates convolutional neural networks (CNNs) and involution neural networks (INNs) to address these issues. CNNs are utilized for capturing global frequency structures, while INNs adaptively extract localized spatial features, significantly improving instrument recognition. The ensemble learning strategy, leveraging soft voting, further enhances the robustness and generalization of the model by aggregating predictions from both CNNs and INNs. This approach directly processes variable-length polyphonic audio, reducing computational complexity while improving efficiency and accuracy. The article provides a comprehensive evaluation using the IRMAS dataset, demonstrating superior performance in both single and multiple predominant instrument recognition tasks. The proposed method offers a scalable and efficient solution for automated music retrieval and classification, making it a significant advancement in the field of music information retrieval.