RETRACTED: Rigdelet neural network and improved partial reinforcement effect optimizer for music genre classification from sound spectrum images

In this paper, a new approach has been introduced for classifying the music genres. The proposed approach involves transforming an audio signal into a unified representation known as a sound spectrum, from which texture features have been extracted using an enhanced Rigdelet Neural Network (RNN). Additionally, the RNN has been optimized using an improved version of the partial reinforcement effect optimizer (IPREO) that effectively avoids local optima and enhances the RNN's generalization capability. The GTZAN dataset has been utilized in experiments to assess the effectiveness of the proposed RNN/IPREO model for music genre classification. The results show an impressive accuracy of 92 % by incorporating a combination of spectral centroid, Mel-spectrogram, and Mel-frequency cepstral coefficients (MFCCs) as features. This performance significantly outperformed K-Means (58 %) and Support Vector Machines (up to 68 %). Furthermore, the RNN/IPREO model outshined various deep learning architectures such as Neural Networks (65 %), RNNs (84 %), CNNs (88 %), DNNs (86 %), VGG-16 (91 %), and ResNet-50 (90 %). It is worth noting that the RNN/IPREO model was able to achieve comparable results to well-known deep models like VGG-16, ResNet-50, and RNN-LSTM, sometimes even surpassing their scores. This highlights the strength of its hybrid CNN-Bi-directional RNN design in conjunction with the IPREO parameter optimization algorithm for extracting intricate and sequential auditory data.