Optimized Cascaded Deep Capsule Cell Attention Network for Efficient Music Genre Classification

This research work develops a new approach towards music genre prediction by fusing high-end signal processing and machine learning methodologies that will improve prediction quality and resilience. The system commences by signal preprocessing comprising Discrete Wavelet Transformation (DWT) in combination with Pre-Gaussian removal of noise. Feature extraction is done with the Mellin Transform that extracts the important features, which immune to changes in pitch and variations in the general characteristics of the different genre of a particular artist or singer. Hierarchical classification is realized by the proposed Cascaded Capsule Cell Attention Network (CCCAN), which utilizes the capsule networks’ advantage of dynamic routing and the spatial hierarchy to capture and identify the intricate structures of audio patterns. The parameter of the network is optimized with Boosted Sooty Tern Optimization so that convergence speed and classification accuracy is above par. Experimental results presented in this study reveal that the proposed model provides around 99.5% of accuracy, precision, recall, and F1-score improved in comparison to the previous method which provides reliable classification of music genres.