基于双向椭圆局部二值模式的环境声音分类

Ambient sound classification based on bidirectional elliptic local binary pattern

针对目前机器学习算法在环境声音分类准确率不高,训练速度慢的问题,提出了基于双向椭圆局部二值模式的环境声音分类方法。设计了双向椭圆局部二进制模式的音频信号特征提取方法,采用3×5信号邻域增加时长影响,并使用邻域左右两列整体平均值分别代替椭圆左右顶点像素,减少噪音干扰,提高对噪音的鲁棒性,使用整个邻域的平均值代替中心像素,并采用双向局部特征均衡顺序权重,在上述特征基础上增加VAR算子,反应局部特征差异强度,之后将这些特征与梅尔频率倒谱系数(MFCC)、伽玛频率倒谱系数(GFCC)和色度特征(Chromagram)融合。采用经典机器学习算法,如支持向量机(SVM)、随机森林(RF)和k近邻(kNN),结合融合特征,在ESC-10和ESC-50数据集上进行评估,两种数据集的分类准确度分别达到了90.9%和66.7%。

Aiming at the problems of low accuracy and slow training speed of current machine learning algorithms in ambient sound classification,a new ambient sound classification method based on bidirectional elliptical local binary mode is proposed.Design a feature extraction method for audio signals in bidirectional elliptical local binary mode,to increase the impact of time length,a 3×5 signal neighborhood is used,and uses the overall average of the left and right columns of the neighborhood to replace the left and right vertex pixels of the ellipse,reducing noise interference and improving robustness to noise,using the average value of the entire neighborhood instead of the center pixel,uses bidirectional local features to balance sequential weights,add VAR operators to the above features to reflect the difference strength of local features,and then fuse these features with MFCC,GFCC,and Chrome features.Using classical machine learning algorithms(support vector machines(SVM),random forests,and k-nearest neighbors(kNN))combined with fusion features,the classification accuracy of the two datasets was 90.9%and 66.7%,respectively,when evaluated on the ESC-10 and ESC-50 datasets.