基于RF-CNN-CBAM-BiLSTM的兴山民歌分类研究

A Study on the Classification of Xingshan Folk Songs Using RF-CNN-CBAM-BiLSTM

摘要: 本研究构建了RF-CNN-CBAM-BiLSTM算法，并对兴山民歌进行分类识别。该模型先利用随机森林(Random Forest, RF)降维方法替代原始的经验选择来挑选训练所用特征，再将卷积块注意力模块(Convolutional Block Attention Module, CBAM)模块融入卷积神经网络(Convolutional Neural Network, CNN)架构，增强模型特征关注与识别能力，接着利用双向长短期记忆网络(Bidirectional Long Short-Term Memory, BiLSTM)捕捉音频序列双向上下文信息，提升兴山民歌的识别性能。本研究所提模型的识别准确率达到91.67%，每一轮的运行时间为22 s，其准确率超过残差网络(Residual Network 50, ResNet50)，高效网络(Rethinking Model Scaling for Convolutional Neural Networks, EfficientNetV1)等四种基准模型的平均准确率，约5.24%。每一轮运行速度较四种基准模型的平均运行时间缩短68.25 s。

Abstract: This study proposes an audio classification algorithm based on the RF-CNN-CBAM-BiLSTM architecture for the classification and recognition of Xingshan folk songs. The model first employs Random Forest (RF) for feature dimensionality reduction, replacing traditional empirical feature selection. Then, the Convolutional Block Attention Module (CBAM) is integrated into the Convolutional Neural Network (CNN) architecture to enhance the model’s ability to focus on and extract relevant features. Subsequently, a Bidirectional Long Short-Term Memory (BiLSTM) network is utilized to capture bidirectional contextual information in audio sequences, further improving the recognition performance of Xingshan folk songs. The proposed model achieves a recognition accuracy of 91.67%, with an average runtime of 22 seconds per training epoch. Its accuracy surpasses the average performance of four benchmark models, including the Residual Network (ResNet50) and EfficientNetV1, by approximately 5.24%. Additionally, the average runtime per epoch is reduced by 68.25 seconds compared to these benchmarks.