Lightweight Deep-Learning Based Music Genre Classification: A Study

Deep-learning (DL) applications that are used real-time across various industries have gained a lot of traction and have become increasingly popular, especially when it comes to data-driven recommendation systems. This work aims to develop a DL scheme to support the music-recommendation system (MS) based on the music data. The various phases of this scheme includes; (i) data collection and signal-image conversion to get the necessary RGB scale images from the data, (ii) pre-trained DL based feature extraction, and (iii) deep-features based detection to recommend the appropriate music. This research considered the classic- (CL) and pop-music (PO) for the examination and the achieved results are evaluated to substantiate the performance of this arrangement. In this work, the signal-image conversion procedure is implemented to convert 1D signal to 2D image and then it is examined using proposed technique. The experimental outcome is separately presented for (i) spectrogram and (ii) synchro-extracting-transform and obtained results are presented. The experimental investigation is presented with MobileNet variants and this study authorizes that the implemented scheme achieved a better detection MobileNetV2 (>99%) compared to other schemes in this study.