You are on page 1of 4

Hybrid Noise Reduction for Audio Captured by Drones

Hyohoon Ahn Duc Tai Le Thien Binh Dang


Sungkyunkwan University Sungkyunkwan University Sungkyunkwan University
Suwon, South Korea Suwon, South Korea Suwon, South Korea
hyohoon@skku.edu ldtai@skku.edu dtbinh@skku.edu

Siwon Kim Hyunseung Choo*


Sungkyunkwan University College of Software
Suwon, South Korea Sungkyunkwan University
siwon.kim59@gmail.com Suwon, South Korea
choo@skku.edu

ABSTRACT ACM Reference format:

Recently, advances in technology have reduced the sizes, lowered H. Ahn, D. T. Le, T. B. Dang, S. Kim, and H. Choo. 2018. Hybrid Noise
Reduction for Audio Captured by Drones, In Proceedings of 12th
prices, and increased commercial use of drones. In the market,
International conference on Ubiquitous Information Management and
drones have become an inexpensive and accessible way to communication (IMCOM’18), Langkawi, Malaysia, January 2018, 4 pages.
accomplish many tasks such as filming, surveillance, agriculture, DOI: 10.1145/3164541.3164576
delivery, and broadcasting. Such applications typically require a
video recording for their operations. However, due to the noise
1 INTRODUCTION
generated by propellers and rotors of a drone, the quality of sound
in a video recorded by the drone is quite limited. In order to solve In recent years, advancement in technology have resulted in size
the noise problem of drones, this paper proposes a novel hybrid reduction, low prices, and increased commercial and personal use
noise reduction technique combining Active Noise Control and of drones [1]. In the market, drones are used for various
Spectral subtraction methods. Experimental results show that the applications and services such as filming, surveillance, agriculture,
proposed technique produces a speech signal with up to 67.5% delivery, and broadcasting. Most of these applications and services
similarity to the original signal. It outperforms Active Noise require video recording for their operations.
Control and Spectral Subtraction with similarities of 53.1% and The first generation drones were difficult to use due to their
39.6%, respectively. complex of hardware and software configuration and flight
manipulation. Furthermore, being expensive and large size, they
CCS CONCEPTS were not suitable for general purposes. However, as the technology
developed, the size of the drone became smaller and the price
• Hardware → Digital signal processing; • Information systems
became lower, the accessibility has increased. In addition, the drone
→ Information systems applications;
manufacturers provide various applications and many utilizing
functions are opened as open sources. Those applications and open
KEYWORDS sources help users to capture images with stable flight and allow
Internet of Things, Drones, Noise reduction, Active Noise users to achieve various missions.
Cancellation, Spectral subtraction Aerial photography using drones can extend the range of view
compared to ground photography by showing different scenes from
the sky. A flying drone provides multiple video capturing options
* Corresponding Author
for filming videos. However, due to the noise generated by
propellers and rotors of a drone, the quality of sound in a video
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed captured by the drone is limited [2]. The noise of propellers and
for profit or commercial advantage and that copies bear this notice and the full citation rotors of the drone makes it difficult to utilize the audio in a video
on the first page. Copyrights for components of this work owned by others than ACM recorded by the drone. In a film and a broadcast, the video and the
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a sound can be recorded separately by using external microphones.
fee. Request permissions from Permissions@acm.org. However, there is a limitation in separating the video and sound
IMCOM '18, January 5–7, 2018, Langkawi, Malaysia using an external microphone if the sound is processed with the
Copyright 2018 Association for Computing Machinery.
video in real time, such as a scene at a news site or a travel
ACM ISBN 978-1-4503-6385-3/18/01…$15.00
https://doi.org/10.1145/3164541.3175682 destination, or a realistic feeling is required for the video.
IMCOM’18, January 5-7, Langkawi, Malaysia H. Ahn et al.

In order to solve the noise problem of drones, this paper proposes size and weight of the object compared to the PNC. Fig. 1 shows
a novel hybrid noise reduction technique that combines Active the operating principle of ANC. To reduce noise, the ANC detects
Noise Control and Spectral subtraction to reduce noise of drones. background noise through an additional microphone. The ANC
In addition, we demonstrate the superiority of the proposed method uses the background noise to generate the opposite sound wave and
by comparing the experimental results of the Hybrid noise emit it through the speaker to control the noise around the speaker.
reduction with the results of existing noise reduction methods. The existing noise and the opposite sound wave emitted through
The rest of the paper is organized as follows. Section 2 the speaker cause interference and cancel to reduce the noise.
introduces Passive Noise Control (PNC), Active Noise Control However, in the real environment, it is difficult to reduce noise by
(ANC), and Spectral subtraction. Section 3 proposes Hybrid noise simply emitting opposing waves because superimposed sound
reduction. Section 4 shows the experimental results using ANC, waves should be exact opposite to cancel noise. The ANC uses an
PNC, and Hybrid Noise Reduction. Section 5 presents the adaptive filter to control the changing noise in the real environment.
conclusion and the future work. An adaptive filter is a filter that operates in a direction that can
obtain optimum results with the ability to adapt its characteristics
2 RELATED WORK according to a change in an input signal or a change in an output.
The adaptive filter does not set all the characteristics at the time of
The background noise is the most common factor worsening the
designing the filter and the characteristics are determined according
quality of audio in recordings. The noise reduction methods intend
to the environment after the operation of the filter. The most
to lower the noise level without affecting the desired audio signal
commonly used adaptive filter is the Least Mean Square (LMS)
quality. These noise reduction methods are based on the Passive
filter algorithm [4-6]. The LMS filter removes the noise signal from
Noise Control (PNC), Active Noise Control (ANC), and Spectral
the received signal by detecting and minimizing the relative error
subtraction.
between the desired signal and the noise signal. The LMS algorithm
PNC is a noise reduction method using the sound absorbing
aims to find an adaptive filter with a minimum error using the
material and sound insulating material. A sound absorbing material
weight setting [7].
absorbs part of the sound energy that hits the surface of the sound
Spectral subtraction method is a single-input noise reduction
absorbing material. The absorber can reduce the reflected sound
method developed for speech enhancement by reducing the white
and the noise. The sound insulating material is installed as an
noise and restoring the power spectrum of the observed signal [8,
obstacle in the path of the sound waves, and reflects the sound
10]. In this method, an average signal spectrum and average noise
waves hitting the wall to reduce the energy of the sound waves to
spectrum are estimated in parts of the recording. These spectrums
be transmitted. PNC is effective for high-frequency and it is easy
are subtracted from each other to improve average signal-to-noise
to apply, but due to the price and weight of the soundproofing
ratio (SNR). SNR is a measure that compares the level of a desired
material, there is a limitation applying it to the air vehicle [3]. In
signal to the level of background noise. Noise estimation is
order to overcome these limitations, researchers have developed an
important to get the effective result of the spectral subtraction [9].
Active Noise Control method.
It is assumed that the noise estimate is the same during the
calculation and the restoration. Also, the phase is the same in the
original signal and restored signal. Noise is estimated using the
period of signal where only noise exists without any desired signal.
Once the noise estimation is calculated, noise spectrum is
subtracted to the input magnitude spectrum or power spectrum.
Spectral subtraction removes noise by using the power spectrum of
sound, so it can’t remove noise if the noise to be removed is larger
than the desired signal to enhance.

3 HYBRID NOISE REDUCTION


In this paper, Hybrid noise reduction is proposed to solve noise
problem in recordings captured by drones. Hybrid noise reduction
is a combination of ANC and Spectral subtraction. The noise of
Figure 1: Active Noise Control propellers and rotors captured by a drone shows a constant
ANC is a noise reduction method that detects disturbed noises waveform pattern. To remove the noise using the concept of ANC
and reduces noise by generating another noise that is opposite to and LMS adaptive filter, it is necessary to use pure drone noise as
the other noise. In ANC, interference between two noise signals an input to generate inverse noise to cancel noise from a target
occurs because sound is transmitted as waves. Cancellation during sound signal. In case of a drone, a constant waveform pattern which
interference is a phenomenon in which mutually opposing waves drone noise shows makes it easier to detect the noise in a target
are mutually extinguished when superimposed. ANC reduces noise sound signal. To reduce the noise from the target sound, the
using destructive interference. The ANC deals effectively with the proposed technique uses two microphones attached to the drone.
The first microphone is for recording a target sound which contains
Hybrid noise reduction for audio captured by drones IMCOM’18, January 2018, Langkawi, Malaysia

desired sound, such as speech, and noise of propellers and rotors. signal that contains propeller noise captured by a drone and is
The second microphone is for recording noise of propellers and recorded from the second microphone of the drone. Noisy speech
rotors only. is a signal that is a combination of speech and drone noise and is
the target signal recorded from the first microphone of the drone.

Figure 2: System Architecture

The operation of the proposed technique is as follows. The


drone records sound from the target. While recording, the drone
uses two microphones to simultaneously acquire the target sound
signal and the noise only signal. The drone applies ANC with LMS
Figure 3: Noise reduction with ANC.
adaptive filter using two audio signals, desired and noise, captured
by the drone as input signals. The noise only signal is used to
In noisy speech, it is difficult to find similarity with speech signal
generate a sound wave that has opposite waveform to the noise due to its noise. Noise reduced signal can be obtained by applying
signal in the target signal. The drone emits generated inverse noise ANC to the noisy speech signal using drone noise as a secondary
through the speaker to control the noise around the drone. The input. The noise reduced signal and the original speech signal
primary noise control allows the drone to record audio which has obtained from the experiment show similarity of 53.1%.
lower lever of noise than desired signal. After the primary noise
reduction is processed and noise reduced audio is recorded, the 4.2 Noise reduction with Spectral subtraction
secondary noise reduction with Spectral subtraction is processed to Spectral subtraction is a single-input noise reduction technique
reduce the remaining noise. The noise detected in the result of which requires only target sound signal that contains a desired
primary noise reduction has weaker power spectrum than desired signal and a noise. Fig. 4 shows the experimental results obtained
signal. Spectral subtraction can achieve efficient result when a by applying spectral subtraction to the noise captured by the drone.
noise to be reduced has smaller magnitude or power spectrum. As
a result of the secondary noise reduction, clean audio can be
obtained.

4 EXPERIMENTS
We have experimented noise reduction with three technique, ANC,
Spectral subtraction, and Hybrid noise reduction, to compare
performance of each techniques. Experiments were carried out
using recorded speech signals of the person and the noise of
propellers and rotors captured by hovering drone.. The
experimental method is as follows. Combining the speech signal
and the noise signal, a target signal is generated. It is typical form
of sound recorded by the drones which is mixed with the noise.

4.1 Noise reduction with ANC


ANC is double-input noise reduction technique which requires
target sound signal, which contains desired signal and noise, and
noise only signal. Noise only signal is used to generate inverse Figure 4: Noise reduction with Spectral subtraction.
noise to cancel noise from target sound signal. Fig. 3 shows the
experimental results with ANC. In Fig. 3, speech is the original In Fig. 4, speech is the recorded voice of the speaker. Drone noise
speech sound recorded from the person. Drone noise is a noise only is a noise signal from a drone and is recorded from the second

3
IMCOM’18, January 5-7, Langkawi, Malaysia H. Ahn et al.

microphone of the drone. Noisy speech is a signal that is a show 67.5% similarity. Experimental results show that Hybrid
combination of speech and drone noise and is the target signal noise reduction is effective for noise reduction in drones than
recorded from the first microphone of the drone. Spectral conventional noise reduction methods.
subtraction uses only one input signal unlike ANC. Spectral
subtraction uses Noisy speech to detect noise only period and
generates predictive noise. Predictive noise is used for subtracting 5 CONCLUSION
noise. Noise reduced signal is the result of reducing noise by In this paper, we propose Hybrid Noise reduction, a noise control
applying spectral subtraction to the noisy speech signal. The noise method combining ANC and spectral subtraction to solve the noise
reduced signal and the original speech signal obtained from the problem of drones. Proposed method has two phases in noise
experiment show similarity of 39.6%. Due to the characteristic of reduction. First phase is noise reduction in the recording moment
spectral subtraction which it can’t control noise effectively when
of the drone. In the first phase, ANC is applied to control the
noise is greater than desired signal, the result obtained with spectral
primary noise of the drone. After the first phase, the spectral
subtraction is worse than the result obtained with ANC under same
experiment condition. subtraction is applied to the result of the first phase to reduce
remaining noise in the audio obtained from ANC. Experimental
4.3 Noise reduction with Hybrid noise reduction result shows that the noise reduced by Hybrid Noise Reduction has
67.5% similarity compared to the original audio. The experimental
Hybrid noise reduction is a combination of ANC and spectral
results of ANC and spectral subtraction show similarity of 53.1%
subtraction. It has two phases of noise reduction. In first phase,
and 39.6% to the original speech signal. The results show that the
primary noise reduction is processed with ANC while the drone is
proposed noise reduction method can operate more effectively in
recording a target sound. A noise only signal and a target sound
the drone recording environment than the conventional noise
signal are necessary for applying ANC in this phase. Same as the
reducing method.
experiment with ANC, the drone generates inverse noise signal
using noise input and emit it through speaker. Two noises with
opposite waveform cancel each other causing interference. As a ACKNOWLEDGMENTS
result of the first phase, a drone can capture primary noise reduced This research was partially supported by PRCP (NRF-2010-
target sound. After recording noise reduced signal by ANC, the 0020210) and G-ITRC support program (IITP-2017-2015-0-
secondary noise reduction phase is processed with spectral 00742), respectively.
subtraction. The second phase uses noise reduced target sound
obtained by the first phase as an input signal. Spectral subtraction REFERENCES
reduces the remaining noise from the audio obtained by ANC. [1] L. Gupta, R. Jain and G. Vaszkun, "Survey of Important Issues in UAV
Communication Networks," IEEE Communications Surveys & Tutorials,
vol. 18, no. 2, pp. 1123-1152, 2016.
[2] G. Sinibaldi, and L. Marino, "Experimental analysis on the noise of
propellers for small UAV," Applied Acoustics, 74, no. 1, pp. 79-88, 2013.
[3] S. J. Elliott and P. A. Nelson, "Active noise control," IEEE Signal Processing
Magazine, vol. 10, no. 4, pp. 12-35, Oct. 1993.
[4] S. M. Kuo and D. R. Morgan, "Active noise control: a tutorial review,"
Proceedings of the IEEE, vol. 87, no. 6, pp. 943-973, Jun 1999.
[5] P. N. Samarasinghe, W. Zhang and T. D. Abhayapala, "Recent Advances in
Active Noise Control Inside Automobile Cabins: Toward quieter cars," IEEE
Signal Processing Magazine, vol. 33, no. 6, pp. 61-73, Nov. 2016.
[6] I. T. Ardekani and W. H. Abdulla, "Active Noise Control in Three
Dimensions," IEEE Transactions on Control Systems Technology, vol. 22,
no. 6, pp. 2150-2159, Nov. 2014.
[7] N. V. George, and G. Panda, “Advances in active noise control: A survey,
with emphasis on recent nonlinear techniques,” Signal processing, Vol. 93,
pp. 363-377, 2013.
[8] M. A. A. El-Fattah, M. I., Dessouky, A. M., Abbas, S. M. , Diab, E. S. M. ,
El-Rabaie, W. , Al-Nuaimy, and F. E. A. El-Samie, “Speech enhancement
with an adaptive Wiener filter,” International Journal of Speech Technology,
Vol. 17, pp. 53-64, 2014.
[9] H. T. Hu and C. Yu, "Adaptive noise spectral estimation for spectral
subtraction speech enhancement," IET Signal Processing, vol. 1, no. 3, pp.
156-163, 2007.
[10] S. Boll, "Suppression of acoustic noise in speech using spectral subtraction,"
IEEE Transactions on Acoustics, Speech, and Signal Processing, vol. 27, no.
2, pp. 113-120, 1979.
Figure 5: Noise reduction with Hybrid Noise Reduction

Fig. 5 shows the result of applying hybrid noise reduction to drones.


Speech is an original speech of a person without any noise. Result
of ANC is same signal as Noise reduced signal in Fig. 3. Noise
reduced signal in Fig. 5 is the result of hybrid noise reduction by
applying Spectral subtraction to result of ANC. The noise reduced
signal and the original speech signal obtained from the experiment

You might also like