基于Vision Transformer和迁移学习的家庭领域哭声识别  

Classification of infant cry sounds based on Vision Transformer and transfer learning

作　　者：王汝旭 王荣燕[1] 曾科 杨传德 刘超[1] WANG Ruxu;WANG Rongyan;ZENG Ke;YANG Chuande;LIU Chao(School of Computer and Information,Dezhou University,Dezhou 253023,Shandong,China)

机构地区：[1]德州学院计算机与信息学院,山东德州253023

出　　处：《智能计算机与应用》2024年第6期119-126,共8页Intelligent Computer and Applications

基　　金：国家级大学生创新训练项目(202210448014)。

摘　　要：针对SVM等传统机器学习算法准确率低和当前使用CNN处理家庭领域哭声识别在不同婴儿间出现泛化能力差的问题,提出了一种基于Vision Transformer和迁移学习的婴儿哭声音频分类算法。首先,为实现数据集样本的扩增,采用了包括梅尔频谱转换和数据增强的数据预处理技术,进而达到了增强模型鲁棒性的目的。而后,在微调后的Vision Transformer模型上进行迁移学习训练,同时,训练过程中利用了LookAhead优化器来不断调整模型参数以避免过拟合,最终实验实现了对婴儿哭声音频的自动分类。实验结果表明,本实验模型相比其他深度学习模型具有更高的精确率和更快的收敛速度,同时还能有效地学习到婴儿哭声中更具区分性的特征。可以在新生儿监护、听力筛查和异常检测等领域中发挥重要作用。Aiming at the low accuracy of traditional machine learning algorithms such as SVM and the poor generalization ability of the current CNN in dealing with cry recognition in the family field between different infants,the infant cry audio classification algorithm based on Vision Transformer and transfer learning is proposed.Firstly,in order to realize the expansion of the data set samples,the data preprocessing technology including MEL spectrum conversion and data augmentation is used,so as to achieve the purpose of enhancing the robustness of the model.Then,transfer learning training is performed on the fine-tuned Vision Transformer model.At the same time,the LookAhead optimizer is used to continuously adjust the model parameters in the training process to avoid overfitting.Finally,the research realizes the automatic classification of infant crying audio.The experimental results show that the proposed model has higher accuracy and faster convergence speed than other deep learning models,and can effectively learn more discriminative features in infant crying.The research can play an important role in the fields of neonatal monitoring,hearing screening and anomaly detection.

关 键 词：Vision Transformer模型 婴儿哭声 迁移学习 梅尔频谱图 LOOKAHEAD 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]