基于动物鸣声的物种识别研究
文献类型:学位论文
作者 | 李娜 |
学位类别 | 硕士 |
答辩日期 | 2013-05 |
授予单位 | 中国科学院研究生院 |
授予地点 | 北京 |
导师 | 朱建国 |
关键词 | 鸣声识别 梅尔倒谱系数 随机森林 |
其他题名 | Idengtify Wildlife Species by Their Sounds |
学位专业 | 动物学 |
中文摘要 | 动物鸣声识别,即借助于动物的叫声对发声对象进行物种、个体、性别甚至音节类型的识别。动物鸣声识别在野外监测中的无干扰性和低耗费特点,激发了生态学、动物学、行为学研究工作者对其相关技术的探索。近二十年来,借鉴于语音识别技术和说话人识别技术,动物鸣声识别领域的技术探索有了较大的发展。 动物鸣声中,鸟类的鸣声最为复杂,一般由音素、音节、片段及一段完整的鸣唱组成。在鸣声识别技术中,音节是鸣声识别的基本单位,然而关于音节的分割方法目前尚无统一的技术标准。鸣声识别一般包括训练数据与检测数据采集,鸣声特征提取,训练和识别四个过程,其中特征提取和训练识别是决定识别准确度的关键技术。在表示鸣声的特征中,梅尔倒谱系数(MFCC)是目前应用最广泛的声音特征。随机森林算法是在进行数据分析的机器学习方法中快速优良的分类器,但目前尚未应用于鸣声识别相关领域。 在对鸣声识别领域发展现状进行了总结的基础上,本项工作主要开展了以下三个方面的研究: (1) 采用MFCC和随机森林算法相结合的方法,对台湾的33种蛙类进行了物种识别探究; (2) 采用MFCC和随机森林算法相结合的方法,对154种鸟类的鸣声进行物种识别探究; (3) 探索了用不同音节类型进行训练和识别对鸟类物种识别准确度的影响。 研究结果表明: (1) 对台湾33种蛙类的物种识别准确度达到97.9%,说明将MFCC与随机森林相结合进行蛙类鸣声物种识别可获得较好的识别准确度; (2) 对154种鸟类的物种识别准确度达到84.0%,说明将MFCC与随机森林相结合进行鸟类鸣声物种识别可获得较好的识别准确度; (3) 音节类型能对识别准确度产生影响:训练样本和识别样本之间的音节相似度越高,识别准确度越高。 本鸣声物种识别领域今后需要进一步比较不同的鸣声识别方法,建立统一的技术标准和技术平台,如建立统一的音节分割方法等;从本论文的研究结果来看,若能着重开展动物鸣声资料收集,建立相对完备的鸣声基础数据库,将鸣声识别技术用于野外未知物种的识别是可行的,这将为分类学、生态学等研究,以及动物多样性监测与保护等提供新的技术支撑。 |
英文摘要 | Animal acoustic recognition means recognize animal species, individual, gender or even sound type through calls. The non-invasive and low-cost advantages of sound recognition compared to traditional ringing monitoring in field study motivate ecologists and behavioral scientists and zoologist to study about the methods of automatic animal acoustic recognition. In recent twenty years, prominent progresses have made in the technologies of acoustic recognition and classification based on the improvement got in human speech recognition and speaker recognition. Bird songs are variable and complex, which usually consist of four levels: notes, syllables, phrases and songs. Syllables are commonly used as the basic recognition unit, wherever there is no coherence method in syllable segmentation. Syllable features extracion and classifier for recognition are significant steps for recognition accuracy. Mel frequency cepstral coefficience(Davis and Mermelstein 1980) is the most effective feature used in acoustic recognition. Randomforest, which is a fast calculator and high-efficiency classfier in machine learning areas, has never been used in acoustic recognition. According to pre-works in animal acoustic recognition field, this work mainly focus on there parts of job contents. The first part is to recognize 33 kinds of frog species using the MFCC and randomforest methods. The second part of the work is about engaging MFCC and the randomforest methods to recognize 154 kinds of bird species, and the third part is to examine if the syllable types have effect on recognition accuracy. The results show that: 1. The MFCCs features and randomForest Classifier are suitable for this research, the accuracy for bird species recognition is 85.12%, for frogs is 97.7%. 2.The difference between syllable types used to train and test will vary bird species recognition accuracies, and the more similar the syllable types used in train process and test process, the higher accuracy will be got. |
语种 | 中文 |
公开日期 | 2013-06-09 |
源URL | [http://159.226.149.42:8088/handle/152453/7461] ![]() |
专题 | 昆明动物研究所_动物生态学研究中心 |
推荐引用方式 GB/T 7714 | 李娜. 基于动物鸣声的物种识别研究[D]. 北京. 中国科学院研究生院. 2013. |
入库方式: OAI收割
来源:昆明动物研究所
浏览0
下载0
收藏0
其他版本
除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。