School of Music and Performing Arts, Mianyang Teachers' College, Sichuan, Mianyang 621000, China.
Comput Intell Neurosci. 2022 Jun 7;2022:8724688. doi: 10.1155/2022/8724688. eCollection 2022.
In order to improve the quality and efficiency of music teaching, we try to automate the teaching of music notation. With the addition of computer vision technology and note recognition algorithms, we improve the generative adversarial network to enhance the recognition accuracy and efficiency of music short scores. We adopt an embedded matching structure based on adversarial neural networks, starting from generators and discriminators, respectively, to unify generators and discriminators from the note input side. Each network layer is then laid out according to a cascade structure to preserve the different layers of note features in each convolutional layer. Residual blocks are then inserted in some network layers to break the symmetry of the network structure and enhance the ability of the adversarial network to acquire note features. To verify the efficiency of our method, we select monophonic spectrum, polyphonic spectrum, and miscellaneous spectrum datasets for validation. The experimental results demonstrate that our method has the best recognition accuracy in the monophonic spectrum and the miscellaneous spectrum, which is better than the machine learning method. In the recognition efficiency of note detail information, our method is more efficient in recognition and outperforms other deep learning methods.
为了提高音乐教学的质量和效率，我们尝试实现音乐记谱法的自动化教学。通过添加计算机视觉技术和音符识别算法，我们改进了生成对抗网络，以提高音乐简谱的识别准确率和效率。我们采用基于对抗神经网络的嵌入式匹配结构，分别从生成器和判别器入手，统一从音符输入侧的生成器和判别器。然后，根据级联结构布置每个网络层，以保留每个卷积层中音符特征的不同层。然后在一些网络层中插入残差块，打破网络结构的对称性，增强对抗网络获取音符特征的能力。为了验证我们方法的效率，我们选择了单声道频谱、复调频谱和杂项频谱数据集进行验证。实验结果表明，我们的方法在单声道频谱和杂项频谱中的识别准确率最高，优于机器学习方法。在音符细节信息的识别效率方面，我们的方法在识别方面更高效，优于其他深度学习方法。
