基于滤波器相似性剪枝的声学场景分类

Acoustic scene classification based on filter similarity pruning

为解决场景分类系统模型复杂度高且现有场景分类剪枝算法容易丢失重要特征信息的问题,提出一种基于滤波器相似性剪枝的声学场景分类方法。以滤波器自身相似性为基础,采用余弦距离和欧氏距离度量距离融合的方法生成相似矩阵判断滤波器重要性并消除冗余滤波器。为提高分类器的泛化性,在网络输出端用决策森林后处理并构建网络。实验结果表明,所提方法能够有效在降低模型复杂度的基础上提升分类精度和系统泛化性,尤其是当消除参数较多时,所提方法特别有利。

To solve the problems that the model of scene classification system is complex and the existing scene classification pruning algorithm is easy to lose important feature information,an acoustic scene classification method based on filter similarity pruning was proposed.Based on the similarity of the filter itself,the cosine distance and Euclidean distance were used to measure the distance fusion method to generate the similarity matrix to judge the importance of the filter and eliminate the redundant filter.To improve the generalization of the classifier,the decision forest post-processing was used at the output of the network and the network was constructed.Experimental results show that the proposed method can effectively improve the classification accuracy and system generalization on the basis of reducing the complexity of the model,especially when there are many elimination parameters,the proposed method is particularly advantageous.