基于纹理特征与随机森林的生态声音识别

摘要

In light of non-stationary noises commonly existed in real environment,we present an ecological sounds recognition method which is based on texture features and random forest.The method first uses a noise estimation-based audio enhancement algorithm,i.e.the short-time spectrum estimation,to carry out the front-end processing on sound signals at input end and obtains the enhanced signal power spectrums.Then according to the texture information in these spectrums’graph it employs the sum and difference histogram to make texture features extraction.Finally,the random forest,which is an ensemble classifier based on the decision tree,is adopted for recognition and classification.In experiment part,two sets of contrast tests are designed,the results show that this method not only has good recognition performance,but is also robust to non-stationary noises.%针对真实环境中普遍存在的非平稳噪声，提出一种基于纹理特征与随机森林的生态声音识别方法。该方法首先使用一种基于噪声估计的音频增强算法，即短时谱估计对输入端声音信号进行前端处理，得到增强信号功率谱；然后根据得到的增强信号功率谱图的纹理信息，采用和差统计法对其进行纹理特征提取；最后，利用基于决策树的组合分类器，即随机森林进行识别和分类。设计了两组对比实验，结果表明该方法不仅有良好的识别性能，而且具有噪声鲁棒性。