优化的匹配追踪用于生态声音识别

发布时间：2019-06-25 11:51

【摘要】：近年来,生态环境的保护受到越来越广泛的关注,通过对自然环境中包含的音频信息进行分析和辨识,可以为环境监测,物种勘察等应用提供数据支持,对于保护自然环境,了解生态现状都具有重要意义。针对实际环境中复杂多变的背景噪声干扰分类识别的问题,本文提出了优化的匹配追踪(MP)信号分解方法和新型的多频带信号重构方法用于生态声音信号重构,并重点研究了低信噪比场景下鲁棒的生态声音识别框架。本文的研究工作主要包括以下几个方面：1)优化的信号稀疏分解。针对信号分解计算量较大的问题,提出基于萤火虫算法(GSO)优化的正交匹配追踪(OMP)。相比于MP算法,OMP算法的改进之处在于提高了收敛速度,缺点在于搜索原子字典和原子正交化操作带来的计算量较大。本文利用GSO算法优化最优原子搜索的过程,实现OMP快速稀疏分解。2)抗噪的信号重构。针对噪声环境下的信号稀疏重构无法有效去除噪声成分的缺点,提出OMP稀疏重构和多频带重构结合的两层信号重构方法。第一层,使用OMP稀疏分解作第一阶段重构,保留前景声音的主体结构。第二层,将前阶段分解的剩余分量按频带划分,根据前景声音和背景噪声的频率分布,对重构信号进行增益补偿,完成第二阶段重构。通过两层信号重构的方法能够有效地自适应去除非平稳的环境噪声,提高了对前景声音的重构精度。3)复合特征提取和基于深信度网(DBN)的分类识别框架。根据对生态声音的时频分布,从时域、频域和时一频域提取多种特征集构造复合特征,能够较好的保留时域和频域信息,并弥补了常用的Mel频率倒谱系数(MFCC)抗噪性能不足的缺点。构建不同深度配置的DBN分类器,基于上述特征建立分类模型,实现对生态声音的识别。本文利用实地采集的鸟叫声、动物叫声和昆虫叫声共3类60种声音样本集构建生态声音库,通过将大自然声音按不同信噪比混入纯净测试数据以模拟真实环境下的生态场景。使用两层信号重构的方法对声音进行重构从而达到去噪的目的。接着,使用基于DBN的声音识别框架进行不同场景和信噪比情况下的对比实验。实验表明,基于OMP稀疏重构和多频带重构结合的两层信号重构方法能够有效地抑制噪声,从而提高识别框架的抗噪性和稳定性。与现有生态声音识别方法相比,本文提出的识别框架对生态声音在不同信噪比下的识别性能有不同程度的改善并且具有较好鲁棒性,适合低信噪比噪声情境下使用。
[Abstract]:In recent years, more and more attention has been paid to the protection of ecological environment. Through the analysis and identification of audio information contained in the natural environment, it can provide data support for environmental monitoring, species survey and other applications. It is of great significance to protect the natural environment and understand the ecological situation. In order to solve the problem of classification and recognition of complex and changeable background noise interference in real environment, this paper proposes an optimized matching tracking (MP) signal decomposition method and a new multi-band signal reconstruction method for ecological sound signal reconstruction, and focuses on the robust ecological sound recognition framework in low signal-to-noise ratio (SNR) scenarios. The research work of this paper mainly includes the following aspects: 1) optimized sparse decomposition of signals. In order to solve the problem of large computational complexity of signal decomposition, an orthogonal matching tracking (OMP). Based on firefly algorithm (GSO) optimization is proposed. Compared with the MP algorithm, the improvement of the OMP algorithm lies in the improvement of the convergence speed, but the disadvantage is that the search atomic dictionary and the atomic orthogonal operation bring a large amount of computation. In this paper, GSO algorithm is used to optimize the process of optimal atomic search, and the fast sparse decomposition of OMP is realized. 2) signal reconstruction of anti-noise. In order to solve the problem that signal sparse reconstruction in noisy environment can not effectively remove noise components, a two-layer signal reconstruction method based on OMP sparse reconstruction and multi-band reconstruction is proposed. In the first layer, OMP sparse decomposition is used to reconstruct the first stage, and the main structure of foreground sound is preserved. In the second layer, the residual components decomposed in the previous stage are divided into frequency bands. According to the frequency distribution of foreground sound and background noise, the gain compensation of the reconstructed signal is carried out, and the second stage reconstruction is completed. The two-layer signal reconstruction method can effectively adaptively remove the non-stationary environmental noise and improve the reconstruction accuracy of foreground sound. 3) Composite feature extraction and classification and recognition framework based on deep reliability network (DBN). According to the time-frequency distribution of ecological sound, a variety of feature sets are extracted from time domain, frequency domain and time-frequency domain to construct composite features, which can preserve the time domain and frequency domain information well, and make up for the shortcomings of the common Mel frequency cepstrum coefficient (MFCC) anti-noise performance. The DBN classifiers with different depth configurations are constructed, and the classification model is established based on the above features to realize the recognition of ecological sound. In this paper, the ecological sound database is constructed by using the 60 kinds of sound samples of bird call, animal call and insect call collected in the field, and the ecological scene in the real environment is simulated by mixing the natural sound into the pure test data according to the different signal-to-noise ratio (SNR). The two-layer signal reconstruction method is used to reconstruct the sound so as to achieve the purpose of denoising. Then, the voice recognition framework based on DBN is used to carry out the comparative experiments under different scenarios and signal-to-noise ratio (SNR). The experimental results show that the two-layer signal reconstruction method based on OMP sparse reconstruction and multi-band reconstruction can effectively suppress noise, thus improving the anti-noise and stability of the recognition framework. Compared with the existing ecological sound recognition methods, the recognition framework proposed in this paper improves the recognition performance of ecological sound in different degrees under different signal-to-noise ratio (SNR) and has good robustness, and is suitable for use in the situation of low signal-to-noise ratio (SNR) noise.
【学位授予单位】：福州大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN912.34

【相似文献】