基于仿生特征模式匹配的海洋动物声信号识别*

针对小数据量的海洋动物声信号混合识别,将声信号同态分析过程中的线性频率转换为Mel频率,得到模拟人耳听觉特性的Mel频率倒谱系数作为声信号的特征。按照声信号所属的物种建立特征模板,使用动态时间规整算法对待识别特征进行分类识别,并对特征库和识别算法进行优化。分别提取了6种鱼类、3种虾类、12种鲸类的Mel频率倒谱系数,为每个物种建立特征模板。分3次对3种、5种、6种鱼类进行识别,分别获得了100%、96.25%、94.68%的识别率。对6种鱼类、3种虾类、12种鲸类共21个物种进行混合识别,总识别率由87.56%提升至优化后的88.96%。实验结果表明,基于Mel频率倒谱系数和动态时间规整算法的海洋动物声信号混合识别能够在小数据量时获得较高的识别率,优化后的特征库和识别算法能够提升识别率。

Sound identification of marine animals based on bionic feature pattern matching

In order to identify the sound of marine animals by using a small amount of data, the linear frequency of the sound was transformed to Mel frequency during the homomorphic analysis, the MFCC(Mel Frequency Cepstrum Coefficient) which simulates the human auditory characteristics was extracted as the feature of the sound. The feature template of species was built; the new features were identified by DTW algorithm; the feature template and algorithm were optimized. The MFCC coefficient of 6 species of fish, 3 species of shrimp and 12 species of whale was extracted to build feature template for each species. The identification rate of 3 species, 5 species and 6 species of fish were 100%, 96.25% and 94.68%. The hybrid identification rate of 21 species including 6 species of fish, 3 species of shrimp and 12 species of whale increased from 87.56% to 88.96% after optimizing. The results show that the hybrid identification based on MFCC and DTW can get a good identification rate by using a small amount of data and the optimized feature template and identification algorithm can improve the identification rate.