【摘要】 为了提高乐音主频识别性能，采用胶囊神经网络用于主频识别，并对胶囊神经网络特征相似计算方法进行改进优化，以增强胶囊神经网络的主频识别适应度。对乐音音符的端点检测与有效分割后采用线性预测倒谱参数法获得乐音主频特征向量。建立基于胶囊神经网络的乐音主频识别模型，并采用动态路由获得稳定的胶囊神经网络结构核心参数。采用余弦相似度对传统的内积计算进行有效改进，优化特征差异判断策略。采用改进的胶囊神经网络算法用于乐音主频识别。试验结果证明，合理设置胶囊神经网络的耦合系数、平衡系数和类别阈值单音集和曲谱连续集均能获得较高的乐音主频识别性能。相比于常用乐音识别算法，该文所提算法能够获得更高的识别准确率和稳定性。更多还原

【Abstract】 In order to improve the performance of music dominant frequency recognition, capsule neural network is used for dominant frequency recognition, and the feature similarity calculation method of capsule neural network is improved and optimized to enhance the adaptability of capsule neural network for dominant frequency recognition. First, music after the endpoint detection and effective segmentation of the notes, the linear predictilve cepstrat coefficient(LPCC)method is used to obtain the dominant frequency feature vector of the notes. Then, the dominant frequency recognition model based on the capsule neural network is established, and the stable core parameters of the capsule neural network structure are obtained by dynamic routing, the cosine similarity is used to effectively improve the traditional inner product calculation and optimize the feature difference judgment strategy. Finally, the improved capsule neural network algorithm is used to identify the dominant frequency of music. Experimental results show that the coupling system of capsule neural network balance coefficient and category threshold, are reasonably set up and single tone set or continuous score set can obtain higher performance of music dominant frequency recognition. Compared with the common music recognition algorithms, this algorithm can obtain higher recognition accuracy and stability.更多还原