异常声音检测中模型压缩算法研究

Research on Model Compression in Abnormal Audio Event Detection

深度学习已经在视觉、语音等领域取得了巨大的成功,随着深度学习性能不断的提升,模型的参数也在不断增加。针对如何在保证模型准确度的同时降低模型大小,使其能够部署在物联网设备上的问题,提出了一种异常声音检测模型与模型压缩算法。异常声音检测模型主要包括端点检测、特征提取、卷积检测模型。针对卷积模型较大的问题,提出了一种模型压缩算法,通过计算节点的输入输出权重之和,裁剪对模型影响较小的节点。实验表明,该异常声音检测模型与压缩算法在保持模型准确度的同时,可以一定程度上降低模型的大小。

Deep learning has achieved great success in visual,voice and other fields.With the continuous improvement of deep learning performance,the parameters of the model are also increasing.Aiming at the problem of how to reduce the size of the model while ensuring the accuracy of the model so that it can be deployed on Internet of things devices,an abnormal sound detec⁃tion model and model compression algorithm are proposed.The abnormal sound detection model mainly includes endpoint detec⁃tion,feature extraction and convolution detection model.In view of the larger problem of convolution model,a model compression al⁃gorithm is proposed.By calculating the sum of input and output weights of nodes,the nodes with less influence on the model are clipped.Experiments show that the model and the compression algorithm can reduce the size of the model to some extent while main⁃taining the accuracy of the model.