What bird is that?
(2020) In Master's Theses in Mathematical Sciences MASM01 20201Mathematical Statistics
- Abstract
- The presence of birds in an ecosystem is often a good indicator of the overall biodiversity. Since birds can be hard to see, their sounds are often used instead to measure their presence. To automatically detect birds the most common method is to use a time-frequency representation together with a convolutional neural network. The most used time-frequency representation is called the spectrogram. An alternative to this is the Wigner-Ville distribution (WVD).
The purpose of this thesis is to investigate if bird classification can be improved if the WVD is used instead of the spectrogram.
The bird sounds were gathered from the website xeno-canto.org. Nine bird species were selected and there were in total 859 samples of bird songs.
... (More) - The presence of birds in an ecosystem is often a good indicator of the overall biodiversity. Since birds can be hard to see, their sounds are often used instead to measure their presence. To automatically detect birds the most common method is to use a time-frequency representation together with a convolutional neural network. The most used time-frequency representation is called the spectrogram. An alternative to this is the Wigner-Ville distribution (WVD).
The purpose of this thesis is to investigate if bird classification can be improved if the WVD is used instead of the spectrogram.
The bird sounds were gathered from the website xeno-canto.org. Nine bird species were selected and there were in total 859 samples of bird songs.
To achieve the purpose, four different methods were used. The first one compared the spectrogram to the WVD. The second one compared the spectrogram to the smoothed pseudo Wigner-Ville distribution (SPWVD). The third one compared the spectrogram to several SPWVD's, performed on shorter sound segments. The last one investigated if a high pass filter could improve the methods.
The WVD and its variations performed worse than the spectrogram for all methods. The best result for the spectrogram was 79\% while the best result for the WVD came from a variant of the SPWVD. Its maximum accuracy was 70\%. The poor performance of the WVD is likely, in part, a result of the high computational requirements for the WVD. As a result of this, much shorter sound segments could be utilised for the WVD compared to the spectrogram. In the future it is likely that the computer power available will far exceed the current availability thus giving the WVD a better chance. (Less) - Popular Abstract
- When researchers are assessing the biodiversity of an area, they often look at the birds. This is because a healthy bird population often means a healthy animal and plant life in general. Since birds can be quite hard to see, researchers often use their song to identify the species and how many of them that are present in an area. This process can be quite time consuming and there is a need for automatic methods.
To automate processes like this, one often uses artificial intelligence and machine learning. One of the more commonly used methods for machine learning is called neural networks. For neural networks to work properly they need good data. When it comes to identifying birds using their sound, the neural networks tend to perform... (More) - When researchers are assessing the biodiversity of an area, they often look at the birds. This is because a healthy bird population often means a healthy animal and plant life in general. Since birds can be quite hard to see, researchers often use their song to identify the species and how many of them that are present in an area. This process can be quite time consuming and there is a need for automatic methods.
To automate processes like this, one often uses artificial intelligence and machine learning. One of the more commonly used methods for machine learning is called neural networks. For neural networks to work properly they need good data. When it comes to identifying birds using their sound, the neural networks tend to perform better if the data contains the pitch of their sounds, i.e. their frequencies, rather than the sounds themself.
To find the frequencies one often uses time-frequency representations. These representations visualises how the frequencies of a sound signal vary with time. There are many kinds of time-frequency representations but the most used version for bird song identification is called the spectrogram. There are however other time-frequency representations, one popular choice is called the Wigner-Ville distribution (WVD). It gives a clearer visualization than the spectrogram but takes longer for the computer to compute which means that shorter sound segments must be used. The purpose of this thesis is to investigate if the WVD can improve the accuracy when identifying birds compared to the spectrogram.
To achieve this result four different methods of the WVD and the spectrogram were analysed. The results of the spectrogram and the WVD were then compared. The best result for the spectrogram was 79\% while the best result for the WVD was 70\%. This means that the WVD could not improve the accuracy. One of the shortcomings of the WVD was the computational burden which meant that shorter sound signals had to be used. In the future, when computer power has increased, it is possible that the WVD could achieve better accuracy than the spectrogram. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9023489
- author
- Wetterlundh, Axel LU
- supervisor
- organization
- alternative title
- An attempt in using the Wigner-Ville distribution to identify bird songs
- course
- MASM01 20201
- year
- 2020
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- TimeFrequencyRepresentations WignerVilleDistribution Spectrogram ConvolutionalNeuralNetworks Birds
- publication/series
- Master's Theses in Mathematical Sciences
- report number
- LUNFMS-3094-2020
- ISSN
- 1404-6342
- other publication id
- 2020:E60
- language
- English
- id
- 9023489
- date added to LUP
- 2020-10-05 13:30:48
- date last changed
- 2021-06-04 17:33:23
@misc{9023489,
abstract = {{The presence of birds in an ecosystem is often a good indicator of the overall biodiversity. Since birds can be hard to see, their sounds are often used instead to measure their presence. To automatically detect birds the most common method is to use a time-frequency representation together with a convolutional neural network. The most used time-frequency representation is called the spectrogram. An alternative to this is the Wigner-Ville distribution (WVD).
The purpose of this thesis is to investigate if bird classification can be improved if the WVD is used instead of the spectrogram.
The bird sounds were gathered from the website xeno-canto.org. Nine bird species were selected and there were in total 859 samples of bird songs.
To achieve the purpose, four different methods were used. The first one compared the spectrogram to the WVD. The second one compared the spectrogram to the smoothed pseudo Wigner-Ville distribution (SPWVD). The third one compared the spectrogram to several SPWVD's, performed on shorter sound segments. The last one investigated if a high pass filter could improve the methods.
The WVD and its variations performed worse than the spectrogram for all methods. The best result for the spectrogram was 79\% while the best result for the WVD came from a variant of the SPWVD. Its maximum accuracy was 70\%. The poor performance of the WVD is likely, in part, a result of the high computational requirements for the WVD. As a result of this, much shorter sound segments could be utilised for the WVD compared to the spectrogram. In the future it is likely that the computer power available will far exceed the current availability thus giving the WVD a better chance.}},
author = {{Wetterlundh, Axel}},
issn = {{1404-6342}},
language = {{eng}},
note = {{Student Paper}},
series = {{Master's Theses in Mathematical Sciences}},
title = {{What bird is that?}},
year = {{2020}},
}