I-AM-Bird: A Deep Learning Approach to Detect Amazonian Bird
Species in Residential Environments
Lucas Ferro Zampar
1 a
and Clay Palmeira da Silva
1,2 b
1
Federal University of Amap
´
a, Macap
´
a, Brazil
2
Center for Sustainable Computing, University of Huddersfield, U.K.
Keywords:
Deep Learning, Object Detection, Faster R-CNN, Bird Species.
Abstract:
The Amazon presents several challenges, such as recognizing and monitoring its birdlife. It is known that bird
records are shared by many bird watchers in citizen science initiatives, including by residents who observe
birds feeding at their home feeders. In this context, the work proposed an approach based on deep learning
to automatically detect species of Amazonian birds that frequent residential feeders. To this end, a data set
consisting of 940 images captured by 3 webcams installed in a residential feeder was collected. In total, 1,836
birds of 5 species were recorded and annotated. Then, we used the dataset to train different configurations
of the Faster R-CNN detector. Considering the IoU threshold at 50%, the best model achieved an mAP of
98.33%, an mean precision of 95.96%, and an mean recall of 98.82%. The results also allow us to drive future
works to develop a monitoring system for these species in a citizen science initiative.
1 INTRODUCTION
The Amazon presents several cultural and scientific
challenges. One of them concerns the process of iden-
tifying and monitoring local birdlife. However, cur-
rent Artificial Intelligence (AI) advances offer new
approaches to gradually reduce these difficulties. One
such approach is the automatic detection of bird
species in images.
Bird records are shared in many citizen science
initiatives. For example, bird watchers can publicly
share images of birds on platforms like (WikiAves,
2023) and (eBird, 2023). In this way, they contribute
to ornithological research that studies the migration of
urban birds, climate change, distribution, and species
conservation (Barbosa et al., 2021).
Sharing records can also occur with the help of
feeders where food is placed to attract birds for ob-
servation (Alexandrino et al., 2022). For exam-
ple, the Feeder Watch project has asked its partici-
pants to identify and count the birds that have visited
their home feeders over 30 years, which has already
contributed to dozens of studies (Bonter and Greig,
2021).
Watching birds feeding in residential gardens also
a
https://orcid.org/0009-0009-5462-5114
b
https://orcid.org/0000-0003-0438-581X
positively impacts the well-being and connection with
the nature of local residents (White et al., 2023). Fur-
thermore, providing their food can stimulate the per-
ception of contribution to the lives of these animals
(Dayer et al., 2019).
Currently, Deep Learning (DL) is widely used
to automate image analysis in computer vision. DL
models, known as neural networks, are composed of
multiple layers and can automatically learn different
representations of the data with which they are trained
(LeCun et al., 2015). This makes DL models suitable
for tasks such as object detection in images.
In this task, the model must locate the objects
of interest in the image using rectangular bound-
ing boxes, in addition to identifying their respective
classes (Zaidi et al., 2022). For example, a detector
could locate birds and classify their species in an im-
age.
The architecture of a modern detection model con-
sists of the backbone network, responsible for ex-
tracting features from the images, and the head that
performs the detections (Bochkovskiy et al., 2020).
These models are generally based on convolutional
neural networks (CNNs).
For example, Faster R-CNN (Ren et al., 2015) is
a model that proposes Regions of Interest (RoI) from
which objects are detected. Other algorithms perform
detections without the proposal of regions, such as
542
Zampar, L. and Palmeira da Silva, C.
I-AM-Bird: A Deep Learning Approach to Detect Amazonian Bird Species in Residential Environments.
DOI: 10.5220/0012464500003636
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 16th International Conference on Agents and Artificial Intelligence (ICAART 2024) - Volume 2, pages 542-548
ISBN: 978-989-758-680-4; ISSN: 2184-433X
Proceedings Copyright © 2024 by SCITEPRESS Science and Technology Publications, Lda.
SSD (Liu et al., 2016) and YOLO (Redmon et al.,
2016).
In this context, Faster R-CNN is classified as a
two-stage detector, while SSD and YOLO are single-
stage. In general, two-stage detectors tend to be more
accurate at the cost of longer inference time (Elgendy,
2020).
In this scenario, our work proposes automatically
detecting Amazonian bird species that visit residen-
tial feeders with the Faster R-CNN model. For train-
ing purposes, we captured images of birds feeding
through webcams installed in the feeder of a resi-
dence in the town of Santana, in the Brazilian state
of Amap
´
a, located in the Amazonian region.
To the best of our knowledge, this is the first work
to propose the detection of Amazonian bird species
from images collected at residential feeders. Addi-
tionally, we also created a dataset of images of these
species annotated for the object detection task.
2 RELATED WORK
In scientific literature, it is possible to find works that
focus exclusively on the classification of bird species.
For example, in the work of (Srijan et al., 2021), dif-
ferent CNN architectures were investigated to classify
260 species. (Huang and Basanta, 2021) also studied
CNNs in classifying 29 endemic species of Tawain.
In both cases, a mobile application was developed.
In the work of (Pinheiro and Soares, 2021), im-
ages of 64 species of birds from the Brazilian state of
Esp
´
ırito Santo were collected, in addition to the train-
ing of the ResNet101 model, which achieved an ac-
curacy of 70.12 % with validation data. The authors
highlighted the difficulty faced by the model in scenes
with multiple birds. Object detection becomes more
suitable for these cases when locating and classifying
several objects in the scene.
In this sense, works dedicated to detecting birds
within the context of preventing accidents or failures
were found. (Alqaysi et al., 2021) studied different
configurations of the YOLOv4 model to detect birds
flying close to wind farms. (Shi et al., 2021) em-
ployed an adapted YOLOv5 model to detect birds fly-
ing around airports.
In both cases, the bird species were not taken into
account. (Qiu et al., 2022) used the YOLOv4-tiny
model to detect 20 species of birds related to faults in
power transmission lines.
Other work has focused on detecting species in
natural scenes. (Mao et al., 2021) used the domain
randomization technique to train the Faster R-CNN
model to detect 2 species found in an ecological park.
(Xiang et al., 2022) already proposed improving the
Faster R-CNN model to detect 10 species on small
scales in another park.
The closest work found collected images from
transmissions carried out by the Cornell Lab of Or-
nithology that show birds feeding (Mirugwe et al.,
2022). From there, they used the images to train dif-
ferent Faster R-CNN and SSD model configurations.
However, the study did not consider the species of
birds, nor did it collect original images.
3 METHODOLOGY
This section will detail the work development stages,
including data acquisition, training, and evaluation of
the models. All codes developed and the dataset col-
lected are publicly available on the authors’ GitHub
by the following link https://github.com/Lucas-Zam
par/amazonian birds detector.git.
We highlight that due to the lack of previous
studies, creating and defining a baseline was neces-
sary. Therefore, we called the baseline the model that
serves as a basis for analysis to be compared and im-
proved later, which will be disclosed in the next sec-
tion.
3.1 Data Aquisition
Food such as seeds and fruits were placed to at-
tract birds to the residence’s feeder. The birds were
recorded feeding by 3 Logitech C270 webcams in-
stalled on the feeder columns. The cameras were
connected to a notebook for capturing and storing
the recordings. The figures 1 and 2 show the feeder
used in the residence and the general bird recording
scheme, respectively.
Figure 1: Residential feeder.
I-AM-Bird: A Deep Learning Approach to Detect Amazonian Bird Species in Residential Environments
543
Figure 2: General recording scheme.
The recordings were carried out between
07/09/2022 and 14/09/2022, mostly from 11am to
5pm. The recordings from days 7 to 13 were man-
ually trimmed to exclude moments in which birds
were absent. The resulting clippings were organized
based on the date of recording and the predominant
species. It is worth noting that the recordings from
the day 14 did not participate in this process. Instead,
they are being reserved for inference.
From the clippings, 940 frames that make up the
images in the dataset were extracted. The images
were uploaded to the RoboFlow platform to be an-
notated. The annotation process consisted of drawing
a bounding box over each bird to locate it and associ-
ating it with its species for classification.
As a result, 1,836 annotations were made in Pas-
cal VOC format. In addition, the work identified 5
species popularly known as orange-fronted yellow-
finch (Sicalis columbiana), shiny cowbird (Molothrus
bonariensis), ground dove (Columbina spp.), blue-
gray tanager (Thraupis episcopus) and palm tanager
(Thraupis palmarum).
Two distinct ground dove species were identified,
but it was decided to group them in the same class.
The species were identified by consulting the citi-
zen science platforms (WikiAves, 2023) and (eBird,
2023), in addition to using a regional bird manual
(d’Affonseca et al., 2012).
3.2 Training
Model training used IceVision, a high-level Python
framework for computer vision tasks. The models
used are pre-trained and made available by IceVision
through the MMDetection project. The training was
divided into two consecutive phases, preliminary and
final.
3.2.1 Preliminary Phase
In the preliminary phase, 30% of the data was pre-
served in a set called partial, corresponding to 282
images and 560 annotations. This decision was made
to define a baseline with less training data.
The need to define a baseline arises due to the lack
of previous studies. Therefore, it is necessary to de-
fine a model that serves as a basis for analysis to be
compared and improved later, which will be done in
the next phase.
During the preliminary phase, 70% of the partial
set was used for training, while the remaining 30%
was used for validation. The purpose is to establish
our baseline training configuration. Therefore, the
preliminary phase is still subdivided into two consec-
utive steps: backbone selection and hyperparameter
tuning.
The objective of the first step was to select the
model backbone with the best performance. To this
end, 8 backbones of the ResNet (He et al., 2016)
and ResNeXt (Xie et al., 2017) types were tested.
All backbones were followed by the Feature Pyramid
Network (FPN) (Lin et al., 2017), which helps detect
objects at multiple scales.
Regarding ResNet backbones, networks with 50
and 101 layers were experimented with. Regarding
the ResNeXt backbones, the networks had 101 layers
but with 32 and 64 convolution blocks. Each back-
bone was pre-trained by MMDetection with a learn-
ing rate schedule of 1x and 2x. Table 1 summarizes
the tried backbones.
Table 1: Backbone configuration.
Backbone Layers Conv. Groups LR Schedule
ResNet 50 - 1x
ResNet 50 - 2x
ResNet 101 - 1x
ResNet 101 - 2x
ResNeXt 101 32x4d 1x
ResNeXt 101 32x4d 2x
ResNeXt 101 64x4d 1x
ResNeXt 101 64x4d 2x
The objective of the second step was to find some
hyperparameter adjustments that would improve the
performance of the previous model. In this way, 13
new models were trained with the selected backbone
but with variations in hyperparameters such as the to-
tal number of epochs, batch size, learning rate, image
size, and presizing.
It is worth highlighting that presizing is a tech-
nique that applies transformations sequentially to the
image, requiring a single interpolation at the end
(Howard and Gugger, 2020). Therefore, the image
must initially be resized to a larger size than the train-
ing size.
At the end of the preliminary phase, the model
trained with the backbone selected in the first stage
and with the best hyperparameter adjustment found
in the second stage was defined as the baseline.
ICAART 2024 - 16th International Conference on Agents and Artificial Intelligence
544
3.2.2 Final Phase
In the final phase, a single definitive model was
trained using the same training configuration as the
baseline. However, all data corresponding to 940 im-
ages and 1,836 annotations was used at this phase.
70% of the data was intended for training, while the
remaining 30% for validation. After training, the per-
formance of the definitive model was compared with
that of the baseline.
3.3 Evaluation
The primary evaluation metric in object detection
is the mean Average Precision (mAP), which corre-
sponds to the arithmetic mean of the areas under the
precision-recall curve for each class (Padilla et al.,
2021). To calculate the mAP, defining a minimum
level of overlap between the actual and predicted
bounding boxes, represented by the Intersection over
Union (IoU) metric is necessary (Elgendy, 2020).
We adopted a minimum IoU of 50% and mAP as
the primary evaluation metric. Precision and recall
metrics were calculated for each species as a refer-
ence. Additionally, the FiftyOne framework, which
has several tools for evaluating the models, was used
to calculate all the metrics.
4 RESULTS
This section will present the most significant results
achieved in our work in each training phase. It is
worth noting that all training was conducted on the
RTX 2060 S GPU with 8 GB of VRAM.
4.1 Preliminary Phase
During the backbone selection step, the original train-
ing configuration was defined with 20 epochs, batch
size equal to 1, learning rate at 10
4
, and image size at
512x512 without using the presizing technique. In Ta-
ble 2, it is possible to check the mAP value achieved
by each backbone with the validation data.
The backbone that led to the best performance was
the ResNeXt 101 FPN 32x4d 1x, reaching mAP equal
to 94.17%. One point to highlight is that all ResNeXt-
type backbones achieved better metrics than ResNet-
type backbones.
During the hyperparameter tuning stage, the per-
formance improvement was noted by increasing the
size of the training image to 812x812 and employ-
ing the presizing technique with initial resizing to
Table 2: Performance in the backbone selection stage.
Backbone mAP
ResNet 50 FPN 1x 87.11 %
ResNet 50 FPN 2x 87.18 %
ResNet 101 FPN 1x 87.23 %
ResNet 101 FPN x 87.67 %
ResNeXt 101 FPN 32x4d 1x 94.17 %
ResNeXt 101 FPN 32x4d 2x 88.70 %
ResNeXt 101 FPN 64x4d 1x 92.80 %
ResNeXt 101 FPN 64x4d 2x 88.74 %
1024x1024. The model trained under these condi-
tions achieved a mAP of 94.59%, then defined as the
baseline. The table 3 presents the precision and recall
values achieved by the baseline for each species.
Table 3: Precision and recall for each baseline species.
Specie Precision Recall
1
orange-fronted
yellow-finch
93.44% 98.28%
2 shiny cowbird 90.00% 92.31%
3 ground dove 96.43% 100.00%
4 blue-gray tanager 85.19% 100.00%
5 palm tanager 84.62% 88.00%
Mean 89.93 % 95.72 %
It is noted that both the lowest precision and the
lowest recall occurred for the palm tanager species.
The most excellent precision was for the ground dove
species, while the greatest recall was for the ground
dove and the blue-gray tanager species. On average,
precision and recall were 89.93% and 95.72%, re-
spectively.
4.2 Final Phase
The definitive model was trained for 20 epochs, with
batch size equal to 1, learning rate at 10
4
, and im-
age size at 812x812 with initial resizing to the size of
1024x1024 during presizing. The performance of the
definitive model is compared with that of the baseline
in the 4 table.
Table 4: Performance between the baseline and the defini-
tive model.
Model mAP
Baseline 94.59%
Definitive model 98.33%
There was a percentage growth of 3.95% in the
mAP ratio, demonstrating that training with more data
benefited the definitive model. In table 5, it is possi-
I-AM-Bird: A Deep Learning Approach to Detect Amazonian Bird Species in Residential Environments
545
ble to check the precision and recall metrics of the
definitive model for each species.
Table 5: Precision and recall for each species of the defini-
tive model.
Specie Precision Recall
1
orange-fronted
yellow-finch
93.44% 98.28%
2 shiny cowbird 97.25% 100.00%
3 ground dove 96.63% 97.73%
4 blue-gray tanager 98.77% 100.00%
5 palm tanager 92.00% 97.87%
Mean 95.96% 98.82%
The definitive model achieved higher precision
and recall values than the baseline for each species,
except for the reduction in recall from 100% to
97.73% concerning the ground dove species. Further-
more, it is notable that the lowest precision still occurs
for the palm tanager species, despite its value having
increased from 84.62% to 92%, reflecting a percent-
age growth of 8.72%.
In general, the definitive model presented an av-
erage percentage gain in precision of 6.7% compared
to the baseline, going from 89.93% to 95.96%. The
mean recall had a more negligible percentage growth
of 3.24%, going from 95.72% to 98.82%.
In the figures 3, 4, 5 and 6 it is possible to check
detections carried out for different species. The im-
ages were extracted from the recordings on day 14
and do not belong to the training or validation data.
The number next to the species name is the confi-
dence level, a probability that reflects how confident
the model is of detection.
It is essential to highlight that the baseline we
defined was necessary due to the lack of any previ-
ous work related to detecting Amazonian bird species
that visit home feeders. The closest work carried out
by (Mirugwe et al., 2022) achieved the best mAP of
98.5% but did not focus on detecting bird species, nor
did it use original images.
Instead, it focused only on detecting birds feeding
in images collected from Cornell Lab video streams.
Since it is the only work focusing on detecting birds
in feeders, it was selected for comparison. It is worth
noting that we use less data and multiple classes re-
lated to each species.
5 CONCLUSION
In this work, we investigated detecting Amazonian
bird species that visit residential feeders using the
Faster R-CNN model. As a result, the work pro-
duced a set of 940 images and 1,836 annotations of 5
Figure 3: Detection of the blue-gray tanager species.
Figure 4: Detection of the palm tanager species.
Figure 5: Detection of the orange-fronted yellow-finch,
ground dove and palm tanager species.
Figure 6: Detection of orange-fronted yellow-finch and
shiny cowbird species.
species known as orange-fronted yellow-finch, shiny
cowbird, ground dove, blue-gray tanager, and palm
tanager. With this data, different configurations of the
ICAART 2024 - 16th International Conference on Agents and Artificial Intelligence
546
Faster R-CNN model were trained in two phases. In
the preliminary phase, a baseline was defined, achiev-
ing an mAP of 94.59%, mean precision of 89.93%,
and mean recall of 95.72%. In the final phase, a
definitive model trained on more data achieved an
mAP of 98.33%, an mean precision of 95.96%, and
an mean recall of 98.82%. Given these results, the
work demonstrated the feasibility of applying a deep-
learning approach to detect the species that visit the
feeder of the residence in question. Therefore, there
is an opportunity to develop future work that seeks to
implement a monitoring system for these species in a
citizen science initiative to study them and contribute
to their preservation.
6 FUTURE WORKS
Given the results achieved, it is possible to visualize
future work. New images will be collected and anno-
tated in the short term at the same feeder to increase
the data set. In this case, the annotation can be par-
tially automated using the definitive model.
Images featuring the highlighted birds will be col-
lected on citizen science platforms in the medium
term. Pre-training models with these images can help
them learn richer characteristics of the species, which
can contribute to increased performance and general-
ization capacity.
A system with the Raspberry PI board capable of
acquiring new images autonomously in other homes
using cloud computing will be developed in the long
term. Furthermore, the feasibility of performing de-
tections locally in an AI-on-the-edge approach will be
studied.
REFERENCES
Alexandrino, E. R., Camboim, T. A., Chaves, F. G., Bovo,
A. A. A., da Silva, M. A. G., da Silva, D. A. M.,
Moss, M., Souza, T. P., de Oliveira Santos, C., de Mat-
tos Brito, C. B., et al. (2022). Which birds are brazil-
ians seeing on urban and non-urban feeders? an anal-
ysis based on a collective online birding. Ornithology
Research, 30(2):104–117.
Alqaysi, H., Fedorov, I., Qureshi, F. Z., and O’Nils, M.
(2021). A temporal boosted yolo-based model for
birds detection around wind farms. Journal of imag-
ing, 7(11):227.
Barbosa, K., Develey, P., Ribeiro, M., and Jahn, A. (2021).
The contribution of citizen science to research on mi-
gratory and urban birds in brazil. ornithol res 29: 1–
11.
Bochkovskiy, A., Wang, C.-Y., and Liao, H.-Y. M. (2020).
Yolov4: Optimal speed and accuracy of object detec-
tion. arXiv preprint arXiv:2004.10934.
Bonter, D. N. and Greig, E. I. (2021). Over 30 years of stan-
dardized bird counts at supplementary feeding stations
in north america: A citizen science data report for
project feederwatch. Frontiers in Ecology and Evo-
lution, 9:619682.
Dayer, A. A., Rosenblatt, C., Bonter, D. N., Faulkner, H.,
Hall, R. J., Hochachka, W. M., Phillips, T. B., and
Hawley, D. M. (2019). Observations at backyard bird
feeders influence the emotions and actions of people
that feed birds. People and Nature, 1(2):138–151.
d’Affonseca, A., Cohn-Haft, M., and Macedo, I. T. d.
(2012). Aves da regi
˜
ao de Manaus. Editora INPA.
eBird (2023). ebird - discover a new world of birding...
https://ebird.org/home. Accessed: 2023-11-20.
Elgendy, M. (2020). Deep learning for vision systems. Man-
ning.
He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep resid-
ual learning for image recognition. In Proceedings of
the IEEE conference on computer vision and pattern
recognition, pages 770–778.
Howard, J. and Gugger, S. (2020). Deep Learning for
Coders with fastai and PyTorch. O’Reilly Media.
Huang, Y.-P. and Basanta, H. (2021). Recognition of en-
demic bird species using deep learning models. Ieee
Access, 9:102975–102984.
LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learn-
ing. nature, 521(7553):436–444.
Lin, T.-Y., Doll
´
ar, P., Girshick, R., He, K., Hariharan, B.,
and Belongie, S. (2017). Feature pyramid networks
for object detection. In Proceedings of the IEEE con-
ference on computer vision and pattern recognition,
pages 2117–2125.
Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S.,
Fu, C.-Y., and Berg, A. C. (2016). Ssd: Single shot
multibox detector. In Computer Vision–ECCV 2016:
14th European Conference, Amsterdam, The Nether-
lands, October 11–14, 2016, Proceedings, Part I 14,
pages 21–37. Springer.
Mao, X., Chow, J. K., Tan, P. S., Liu, K.-f., Wu, J., Su, Z.,
Cheong, Y. H., Ooi, G. L., Pang, C. C., and Wang,
Y.-H. (2021). Domain randomization-enhanced deep
learning models for bird detection. Scientific reports,
11(1):639.
Mirugwe, A., Nyirenda, J., and Dufourq, E. (2022). Au-
tomating bird detection based on webcam captured
images using deep learning. In Proceedings of 43rd
Conference of the South African Insti, volume 85,
pages 62–76.
Padilla, R., Passos, W. L., Dias, T. L., Netto, S. L., and
Da Silva, E. A. (2021). A comparative analysis of ob-
ject detection metrics with a companion open-source
toolkit. Electronics, 10(3):279.
Pinheiro, B. and Soares, L. (2021). Sistema de reconheci-
mento autom
´
atico de p
´
assaros da fauna de linhares-es
utilizando redes neurais convolucionais. In Simp
´
osio
Brasileiro de Automac¸
˜
ao Inteligente-SBAI, volume 1.
Qiu, Z., Zhu, X., Liao, C., Shi, D., Kuang, Y., Li, Y., and
Zhang, Y. (2022). Detection of bird species related to
I-AM-Bird: A Deep Learning Approach to Detect Amazonian Bird Species in Residential Environments
547
transmission line faults based on lightweight convolu-
tional neural network. IET Generation, Transmission
& Distribution, 16(5):869–881.
Redmon, J., Divvala, S., Girshick, R., and Farhadi, A.
(2016). You only look once: Unified, real-time object
detection. In Proceedings of the IEEE conference on
computer vision and pattern recognition, pages 779–
788.
Ren, S., He, K., Girshick, R., and Sun, J. (2015). Faster
r-cnn: Towards real-time object detection with region
proposal networks. Advances in neural information
processing systems, 28.
Shi, X., Hu, J., Lei, X., and Xu, S. (2021). Detection of
flying birds in airport monitoring based on improved
yolov5. In 2021 6th International Conference on In-
telligent Computing and Signal Processing (ICSP),
pages 1446–1451. IEEE.
Srijan, Samriddhi, and Gupta (2021). Mobile application
for bird species identification using transfer learn-
ing. In 2021 IEEE International Conference on Ar-
tificial Intelligence in Engineering and Technology
(IICAIET), pages 1–6. IEEE.
White, M. E., Hamlin, I., Butler, C. W., and Richardson, M.
(2023). The joy of birds: The effect of rating for joy
or counting garden bird species on wellbeing, anxiety,
and nature connection. Urban Ecosystems, pages 1–
11.
WikiAves (2023). Wiki aves - a enciclop
´
edia das aves do
brasil. https://www.wikiaves.com.br/. Accessed:
2023-11-20.
Xiang, W., Song, Z., Zhang, G., and Wu, X. (2022). Birds
detection in natural scenes based on improved faster
rcnn. Applied Sciences, 12(12):6094.
Xie, S., Girshick, R., Doll
´
ar, P., Tu, Z., and He, K. (2017).
Aggregated residual transformations for deep neural
networks. In Proceedings of the IEEE conference on
computer vision and pattern recognition, pages 1492–
1500.
Zaidi, S. S. A., Ansari, M. S., Aslam, A., Kanwal, N., As-
ghar, M., and Lee, B. (2022). A survey of modern
deep learning based object detection models. Digital
Signal Processing, 126:103514.
ICAART 2024 - 16th International Conference on Agents and Artificial Intelligence
548