使用深度学习对驼背的呼叫进行自动分类:对神经架构和声学特征表示进行比较研究
1School of Electrical Engineering, Computing and Mathematical Sciences (EECMS), Curtin University, Kent Street, Bentley, WA 6102, Australia.
这项研究开发了一种自动化系统,用于使用神经网络和Mel光谱图检测驼背,实现高精度. 该系统的性能优于以前的方法,为被动声学监控数据分析提供了强大的解决方案.
科学领域:
- 海洋生物声学 海洋生物声学
- 计算智能是一种计算智能.
- 信号处理 信号处理
背景情况:
- 被动声学监控 (PAM) 生成大量数据集,需要自动化分析.
- 驼背的发声对于种群研究至关重要,但很难分类.
- 现有的分类方法可能缺乏准确性和稳定性.
研究的目的:
- 开发和评估一个自动化的驼背检测系统.
- 为了比较不同神经网络架构和特征表示的性能.
- 优化数据处理和增强,以提高检测准确度.
主要方法:
- 从公共存储库中创建了一个精心策划的驼背音频数据集.
- 应用了数据增强技术来扩展数据集.
- 多个神经网络,包括MobileNetV2和自定义的CNN,使用TensorFlow和Keras进行训练.
- 作为特征表示,使用了Mel光谱图和Mel频 Cepstral 系数 (MFCC).
主要成果:
- 在所有模型中,Mel光谱图始终优于MFCC.
- 预训练的MobileNetV2与mel光谱图实现了99.01%的准确性,99%的精度/回忆,以及0.98MCC.
- 采用Mel光谱图的定制CNN实现了98.92%的准确率和0.75%的假负率.
- 基于MFCC的模型显示出更低的稳定性和更高的虚假负率.
结论:
- 与MFCC相比,Mel光谱图是检测驼背的优越特征表示.
- 开发的神经网络模型,特别是MobileNetV2,在自动识别驼背方面表现出高效率.
- 这项研究为分析PAM数据和推进海洋哺乳动物监测提供了强大的框架.
更多相关视频
09:09Foreign Accent and Forensic Speaker Identification in Voice Lineups: The Influence of Acoustic Features Based on Prosody
Published on: September 27, 2024
08:06Automated Acoustic Dispensing for the Serial Dilution of Peptide Agonists in Potency Determination Assays
Published on: November 10, 2016
相关概念视频
Polymer Classification: Architecture
State Space Representation
Consider an RLC circuit, a...
Control Volume and System Representations
The control volume approach considers a stationary region in space through which fluid flows. This region is bounded by a control surface. For instance, in the case of water...
Graphical Representation of Inequalities
Vector Representation of Complex Numbers
Consider a function defined as the product of the complex factors in the numerator divided by the product of the complex factors in the...
Graphical and Analytic Representation of Sinusoids
The first step is measuring the peak-to-peak value, which is twice the amplitude of the sinusoid. This provides information about the maximum voltage swing of the waveform.
Secondly, the period and angular frequency are determined. The period is the time taken for one complete cycle of the waveform, while...
