|
Published Instituto
Tecnológico Superior Corporativo Edwards Deming. Quito - Ecuador Frequency October - December Vol. 1, No. 31, 2026 Pp 1-16 http://centrosuragraria.com/index.php/revista Dates of receipt Received: July 09, 2026 Approved: September 06,
2026 Corresponding author Creative Commons License Creative Commons License,
Attribution-NonCommercial-ShareAlike 4.0
International.https://creativecommons.org/licenses/by-nc-sa/4.0/deed.es |
Gabriela del Carmen Suárez Lizárraga
Carlos Humberto Hernández López
Ph.D. in Science, PROINPA Foundation, Ph.D. in Educational Technology, Mazatlán Institute of
Technology https://orcid.org/0009-0005-5710-8367 Ph.D. in Aquatic Resources, Mazatlán Institute of Technology https://orcid.org/0000-0002-6938-0502
Keywords: Deep learning, YOLO v8, animal welfare, non-invasive monitoring.
Resumen: La visión artificial y los sistemas inteligentes
representan una tendencia clave para automatizar el monitoreo conductual en la
acuicultura intensiva. Sin embargo, los enfoques comerciales actuales suelen
ser costosos, rígidos y dependientes de la observación humana manual. El
objetivo de este trabajo, fue desarrollar, entrenar y validar el
sistema denominado EBISU, un marco de software no invasivo para
evaluar variables cinemáticas en vivo. Metodología: La investigación
es de tipo tecnológica aplicada; se empleó una muestra experimental de
tilapia (Oreochromis niloticus) en un entorno de recirculación controlado.
Como instrumentos y técnicas se utilizaron cámaras digitales acopladas a
nodos IoT, procesamiento de video en paralelo mediante PyQt5 y la
arquitectura de aprendizaje profundo YOLO para la detección y seguimiento
geométrico cuadro por cuadro. Resultados: Los resultados demostraron
que el sistema procesa flujos de video fluidamente, determinando con precisión
si las velocidades de nado indican parámetros normales o estrés
metabólico. Discusión: Se destaca queEBISU supera la sustracción de fondo
clásica al mitigar eficazmente reflejos lumínicos y turbulencias del
agua. Conclusión: El sistema automatiza con éxito la supervisión biológica
continua, reduciendo el error humano y sentando bases metodológicas para
futuros trabajos en acuicultura de precisión en tiempo real.
Palabras clave: Aprendizaje profundo, YOLO v8, bienestar animal,
monitoreo no invasivo.
Introduction
In the field of aquaculture, computer vision is emerging as a
fundamental pillar for the transition to Aquaculture 4.0, enabling the
non-invasive monitoring of living organisms in intensive production
environments. In this regard, significant advances have been made, particularly
in Asia, where such studies help strengthen food security in those countries
given the region’s predominant culinary culture (Biazi
and Marqués, 2023; Liu et al., 2023; He et al., 2026; Ranjan, 2026).
However, the Mexican industry, at least in the northwestern region, lags
behind in the adoption of these instrumentation technologies. The monitoring of
culture tanks continues to rely on manual and empirical observation, which
introduces critical errors from a control engineering perspective, including
the inability to perform continuous monitoring, since the human eye cannot
process visual signals without interruption, thereby preventing the detection
of transient anomalous behaviors (Liu et al., 2023; Cui et al., 2025).
Consequently, there is a lack of digitization of biological variables due to
the absence of a real-time processing system that translates swimming patterns
into quantitative data (angular velocity, stroke frequency, flow density),
causing a significant delay between the onset of a problem and the activation
of life-support systems (aerators or dosing pumps) and also resulting in the
loss of vital information for the preventive diagnosis of diseases or hypoxic
stress (Cai et al., 2024).
In this regard, it is worth noting that the impetus behind this
particular project stemmed from a real-life catastrophic event that occurred at
the experimental facilities of the Mazatlán Institute of Technology. A power
outage disabled the aeration system in white shrimp tanks. Because the incident
occurred outside of the monitoring window, the drop in dissolved oxygen was not
detected in time, resulting in the suffocation and 100% mortality of the
biomass under study. This could have been prevented had there been a system
capable of triggering an alarm upon detecting changes that jeopardize
production.
In general, the increasing intensification of aquaculture systems in the
region has heightened the need to implement continuous monitoring mechanisms
that ensure the welfare of farmed organisms and prevent production losses. In
practice, fish monitoring continues to rely heavily on visual observation by
human operators, who empirically assess variables such as swimming activity,
feeding response, spatial distribution of the school, and the presence of
abnormal behaviors (He et al., 2026). However, this approach has inherent
limitations related to observer subjectivity, the inability to maintain
constant surveillance, and the difficulty of detecting subtle behavioral
changes in real time (Al-Abri et al., 2025).
In this context, computer vision-based monitoring systems have become
one of the most promising technologies due to their ability to continuously and
noninvasively obtain biological information (Biazi
and Marqués, 2023). The digital transformation of aquaculture has given rise to
the paradigm known as Aquaculture 4.0, which integrates smart sensors, the
Internet of Things (IoT), artificial intelligence, and computer vision systems
to optimize production, reduce economic losses, and improve animal welfare (Firzgerald et al., 2025).
Recent studies in precision aquaculture recognize that behavior is one
of the most sensitive indicators of the physiological state of fish (Cai et
al., 2024; Ahmed and Jeba, 2024; Liu et al., 2023). Changes in swimming speed,
alterations in group cohesion, an increase in erratic trajectories, decreased
motor activity, or changes in spatial interaction patterns often occur in
response to stress, hypoxia, infectious diseases, poor water quality, or
adverse environmental changes (Fitzgerald et al., 2025; Zhang et al., 2025). As
a result, automated behavioral analysis has become one of the most significant
lines of research within Aquaculture 4.0 (Cui et al., 2025).
From this perspective, the present study hypothesizes that swimming patterns can function as a digital
behavioral biosensor capable of providing early indications of changes in the
physiological state of fish before visible clinical signs or mortality events
occur. In other words, variations in the kinematic characteristics of
swimming patterns (speed, spatial dispersion, and locomotor activity)
constitute early indicators of physiological stress in farmed fish and can be
automatically detected using computer vision and deep learning techniques (Li
et al., 2024). This hypothesis is based on the
fact that aquatic organisms respond to environmental disturbances through
observable changes in their locomotor activity, which can be objectively
quantified using computer vision and artificial intelligence techniques (Cui et
al., 2025).
Recent advances in fish behavior recognition using computer vision have
demonstrated that kinematic variables such as speed, acceleration,
trajectories, group dispersion, and movement frequency can be automatically
extracted from video sequences and used to identify behaviors associated with
feeding, stress, disease, and animal welfare (Al-Abri et al., 2025). These
approaches make it possible to transform biological patterns—traditionally
interpreted subjectively—into quantitative indicators suitable for
computational analysis (He et al., 2026).
However, according to Fitzgerald et al., 2025, the adoption of these
technologies remains limited in many production systems, particularly on small-
and medium-scale farms. As a result, critical events associated with decreases
in dissolved oxygen, aeration failures, deteriorating water quality, or
pathological processes may go undetected for prolonged periods, especially
during times without human supervision (Sun et al., 2025).
Given this scenario, there is a need to develop an intelligent
electronic system based on computer vision that continuously monitors fish
swimming patterns and translates these behaviors into objective indicators of
biological risk. Automating this task would make it possible to replace
intermittent monitoring with a continuous observation system capable of
generating early warnings when significant deviations from normal farm behavior
are detected, thereby helping to reduce economic losses, improve animal
welfare, and strengthen the principles of smart and sustainable aquaculture (He
et al., 2026; Cui et al., 2025).
Therefore, given this context, the work presented in this document was
carried out. It consists of the development of an intelligent electronic system
based on computer vision for the automated monitoring of aquaculture tanks,
aimed at strengthening the productivity and sustainability of the aquaculture
sector in the Mazatlán region of Sinaloa—a strategic area due to its importance
to fishing and aquaculture activities in northwestern Mexico. The proposal
integrates digital image processing, animal behavior analysis, and real-time
electronic monitoring, enabling the identification of abnormal patterns in fish
associated with disease, stress, or changes in water quality. Using cameras and
visual analysis algorithms, the system provides continuous monitoring of the
aquaculture operation, reducing reliance on manual supervision and facilitating
more accurate and timely decision-making by the producer. Among the benefits of
this system are reduced economic losses, optimized feed utilization, increased
operational efficiency, and improved aquaculture production conditions.
Similarly, it helps drive technological transformation in the primary sector,
promoting a more competitive, sustainable, and innovative aquaculture model in
line with Industry 4.0 trends as applied to the agri-food sector.
Methodology
The
research was conducted using an experimental approach focused on the design and
validation of a smart electronic system for the automated monitoring of aquatic
organisms using computer vision. The proposed architecture, called EBISU,
integrates deep learning techniques, object tracking, and behavioral analysis
to identify variations in swimming patterns associated with potential states of
physiological stress.
The
methodology was structured into seven stages during the first semester of 2026:
literature review, visual data acquisition, training dataset construction,
automatic organism detection using convolutional neural networks, kinematic
behavior analysis, early warning generation, and results analysis. This
approach follows current trends in Aquaculture 4.0, where computer vision is
used as a non-invasive tool for the continuous assessment of animal welfare (Biazi and Marqués, 2023).
When
this project began, for practical reasons, we considered using stochastic
computer vision approaches. This algorithm operates on the principle of
“background subtraction.” The camera assumes that the color of the tank water
is static; if a drastic change in contrast occurs, the algorithm crops that
cluster of pixels, assuming it is a moving organism (Bowmans, 2014). This
methodology failed when applied to real-world farms. The water in production
contains suspended organic matter (feces, feed) and dense columns of dynamic
bubbles produced by the diffusers. The algorithm becomes confused, flagging the
bubbles as invisible fish (false positives). This noise saturated the
computer’s memory and rendered the speed tracking inoperable.
Recognizing
that background subtraction was a problem, Deep Learning using Convolutional
Neural Networks (CNNs) from the YOLOv8 family was adopted. Convolutional
matrices are immune to noise; they do not look for “moving colors,” but rather
evaluate biomorphic textures (fins, skull structure) that have been previously
trained (Ranjan, 2026). In this way, the AI isolates correct visual detections
and ignores the aeration.
For
experimental validation, tilapia (Oreochromis niloticus) specimens reared
in experimental ponds at the Mazatlán Institute of Technology were used.
Tilapia was selected due to its widespread use in aquaculture systems in the
region and its high resistance to variable rearing conditions. Digital cameras
were installed above the tanks to capture video sequences representative of the
system’s operating conditions. Subsequently, frames were extracted from the
obtained sequences, and the organisms were manually labeled using visual
annotation tools. The creation of labeled datasets is a fundamental step in
systems based on deep learning , as it allows models to learn the distinctive
visual features of the organisms of interest and improve their ability to
generalize under varying environmental conditions (Cui et al., 2025). The
dataset shown in Figure 1 was divided into three independent subsets: training
(70%), validation (20%), and test (10%), following best practices for
developing computer vision models (Li et al., 2024).
Figure
1. Comprehensive
model convergence plots. The decline in “Loss” (error) and the stabilization of
the metric upon completion of iterative training can be observed.
Automatic
fish detection was initially performed using the YOLOv8 architecture, one of
the most widely used object detection neural networks today due to its balance
between accuracy and inference speed (Terven et al., 2023). The selection of
YOLOv8 was based on recent studies reporting superior results in aquaculture
applications related to fish counting, organism tracking, disease detection,
and real-time behavior recognition (Li et al., 2024; He et al., 2026).
During
the training process, the model iteratively adjusted its internal parameters
through backpropagation of error using the training and validation datasets.
Performance was evaluated using standard metrics employed in computer vision,
including precision, recall, F1-score, and confusion matrix (Rivadeneira et
al., 2024).
Once
the organisms were detected in each frame, the coordinates corresponding to the
geometric centroid of each individual were calculated. These coordinates were
used as input for a tracking module responsible for associating the identity of
each fish across consecutive frames. Tracking was performed by calculating
Euclidean distances between successive positions, allowing for the
reconstruction of individual trajectories and the estimation of kinematic
variables associated with locomotor behavior (Cai et al., 2024).
Based
on the obtained trajectories, behavioral indicators such as instantaneous
speed, average speed, and spatial dispersion were calculated. Various studies
have shown that these variables represent sensitive biomarkers of animal
welfare and can be used for the early detection of states of stress, hypoxia,
or disease (Fitzgerald et al., 2025; Sun et al., 2025).
The
central hypothesis of this research posits that swimming patterns function as a
digital behavioral biosensor capable of reflecting physiological alterations
before the onset of visible clinical signs. Based on this premise, an
analytical module was developed to identify deviations from the culture’s
normal behavior. The system continuously evaluates the group’s average speed
and inter-individual variability to detect two main conditions:
Lethargy
or an abnormal decrease in locomotor activity.
Erratic
swimming or an abrupt increase in kinematic dispersion.
Both
patterns have been reported in the literature as behavioral responses
associated with physiological stress, deteriorating water quality, and hypoxia
events (Fitzgerald et al., 2025; Zhang et al., 2025; Sun et al., 2025).
When
behavioral variables exceed predefined thresholds, the system generates
automatic alerts via remote messaging services (the Telegram app was used),
thereby mitigating the risk of communication breakdowns on rural farms. The
integration was designed with network exception handling. If an alert fails due
to a lack of internet connectivity, the core architecture continues to process
the data without “freezing” or causing a critical shutdown. It also
incorporates a mathematical “cooldown” mechanism that prevents the repetition
of identical alerts (to prevent spam on the on-duty biologist’s cell phone).
This mechanism allows for timely notification of those responsible for the
crop, facilitating the implementation of corrective actions before significant
damage to the biomass occurs.
To
synthesize the information into a real-time analytical format, parallel logic
processors (PyQt5’s QThread) were used. This allows a
complex interface to generate a statistics dashboard powered by Matplotlib (see
Figure 4). This submodule compiles a history of 1,800 computational cycles of
the tank in memory. It plots continuous curves that show the fluctuation of V_global relative to the lethargic danger threshold and, in
parallel, displays distribution bars that calculate the overall percentage of
operational time during which the test tank has remained in a normal,
lethargic, or erratic state, thereby providing a massive data science
framework.
Figure
2. Interactive
console for species-specific parametric orchestration
Similarly,
to demonstrate that artificial intelligence eliminated the interferences
present in the initial test algorithm, the graphs resulting from the pilot
training were analyzed. The spatial distribution illustrated in Figure 3 shows
how the test bed forced the network to discern specimens located at all bottom
densities of the tank, thereby preventing biased memorization of photographic
quadrants.
Figure
3. Spatial
distribution of dataset labels showing positional vector densities within the
visual matrix
Furthermore,
the normalized confusion matrix shown in Figure 4 demonstrates perfect
discrimination against the background (noise and bubbles). YOLOv8 refuses to
assign biological identities to gaseous entities in the aeration column.
Figure
4. Normalized
confusion matrix of the CNN algorithm
The
peak of the F1 curve illustrated in Figure 5 reflects an optimal weighted index
at a parametric confidence level of ~0.4 for the model, balancing the avoidance
of false positives against the omission of false negatives.
Figure
5. F1 curve
(F1-Score), determining the optimal balance between Precision (Positive Predictives) and Recall (Sensitivity).
Figure
6 shows the automated extraction of the validation batch, where the algorithm
isolates the instances without being corrupted by extraneous factors.
Figure
6. Automated
extraction of the validation batch.
The
system’s performance was evaluated through tests conducted under real-world
operating conditions. The analysis assessed the system’s ability to detect
organisms, the stability of tracking, the accuracy of behavioral analysis, and
the system’s robustness against complex environmental conditions such as the
presence of bubbles, variations in lighting, and high biomass densities. The
results obtained were compared using performance metrics employed in object
detection and behavior recognition systems, allowing for an evaluation of the
system’s viability as a continuous monitoring tool for aquaculture
applications.
Results
The
results obtained during the training of the YOLOv8 model demonstrated stable
convergence of the loss functions associated with classification, localization,
and object detection. The progressive decrease in loss values observed during
the training iterations indicates that the neural network successfully learned
the distinctive visual features of the organisms present in the dataset,
gradually reducing the prediction error. The evolution of the performance
metrics showed consistent behavior between the training and validation sets,
suggesting adequate generalization ability of the model and a low tendency
toward overfitting. These results are consistent with recent research that
identifies YOLOv8 as one of the most efficient architectures for aquaculture
applications requiring real-time organism detection under variable
environmental conditions (He et al., 2026; Al-Abri et al., 2025).
Furthermore, the confusion matrix
obtained during the validation stage demonstrated a high ability to distinguish
between fish and environmental elements, particularly in the presence of visual
interference caused by aeration bubbles, surface reflections, and suspended
particles in the water. This result represents a significant improvement over
traditional background subtraction methods, which tend to exhibit a
considerable increase in false positives in dynamic aquatic environments
(Terven et al., 2023).
Once the organisms were detected,
the tracking algorithm based on centroids and Euclidean distance—as also noted
in the study by García (2024)—allowed for the identification of individual
organisms to be maintained across consecutive video sequences. The tests
conducted demonstrated that the system was capable of reconstructing continuous
trajectories even in situations of high population density and frequent
interactions between organisms. The stability observed during tracking made it
possible to obtain consistent time series of position and movement, an
essential condition for the reliable estimation of behavioral variables. This can be seen in the images obtained from
the system, which are illustrated in Figure 7.
Figure 7. In-Situ Validation of Continuous
Tracking
|
|
|
In this regard, Fitzgerald et al.
(2025) note that tracking accuracy is one of the most critical factors in
automated animal behavior analysis systems, since association errors can
propagate to subsequent stages of kinematic analysis. Therefore, based on the
reconstructed trajectories, kinematic variables related to the fish’s locomotor
behavior were calculated, including instantaneous speed, average group speed,
and spatial distribution of the school (Cui et al., 2025). These variables were
analyzed continuously to identify deviations from behavior considered normal
within the experimental system.
The results of the performance
metrics (see Table 1) showed that the system was capable of detecting
significant changes in the collective dynamics of the fish population by
identifying patterns consistent with states of hyperactivity and lethargy. The
scientific literature recognizes that these types of behavioral responses are
often associated with physiological stress, decreases in dissolved oxygen,
environmental disturbances, and early stages of disease (Sun et al., 2025;
Zhang et al., 2025).
Table 1. Performance metrics of the EBISU
detection model (YOLOv8)
|
Performance Metric |
Value (%) |
|
Precision |
96.2 |
|
Recall |
94.8 |
|
Mean Average Precision (mAP@0.50) |
98.5 |
|
Optimal F1-score |
0.95 |
The integration of artificial
intelligence-based early warning systems is one of the main areas of
development in smart aquaculture, as it enables the transformation of large
volumes of data into useful information for operational decision-making (Li et al.,
2024).
The results suggest that the
proposed architecture has the potential to be integrated into Aquaculture 4.0
frameworks focused on the smart monitoring of aquatic organisms. The
combination of deep learning-based detection, automated tracking, and behavioral
analysis provides a tool capable of transforming complex biological signals
into objective indicators for decision-making.
Additionally, the modular nature of
the system facilitates its adaptation to other aquaculture species by
retraining the detection models while retaining the tracking and behavioral
analysis algorithms. All kinematics, stress mathematics, telemetry, and underlying
architecture will function without altering the base code, thereby
consolidating a universal multispecies framework. This feature aligns with
current research trends in digital aquaculture, which aim to develop scalable
and reusable platforms for different production scenarios (He et al., 2026; Cui
et al., 2025; Sun et al., 2025).
As a future project, the goal is to
train the system using more advanced technologies—such as YOLO v11—to improve
accuracy and sensitivity, among other parameters that may be of interest. This
will require investing in equipment capable of processing the data optimally.
Conclusions
The system’s ability to
automatically quantify variations in fish movement supports the hypothesis that
swimming patterns can be used as a digital behavioral biosensor. From this
perspective, locomotor behavior ceases to be merely a qualitative observation
made by human operators and becomes a continuous source of quantifiable
information for assessing animal welfare.
One of the main findings of this
research was the feasibility of implementing a continuous monitoring system
without the need for constant human supervision. The system operated
autonomously, processing video sequences in real time, calculating behavioral
indicators, and generating automatic alerts when the analyzed variables
exceeded established thresholds. This result is significant because one of the
most important limitations of traditional aquaculture production systems is
their reliance on intermittent visual inspections conducted by human personnel (Fitzgerald
et al., 2025). Recent research highlights that computer vision systems
constitute an effective alternative for increasing monitoring frequency,
reducing errors associated with observer subjectivity, and improving
responsiveness to critical events (Zhang et al., 2025; Sun et al., 2025).
The core value of this research lies
in designing the software architecture under a unified, generic, and robust
paradigm. EBISU has been demonstrated to function as an adaptable Universal
Framework that allows researchers to completely disregard the specific test
species. Simply compiling and substituting the neural weights will suffice to
monitor a rainbow trout station or crustacean maturation, and the entire
mathematical structure of Strikes, Thresholds, Vectors, visual data science, and
HTTP telemetry will function properly.
This orchestration replaces the
vulnerability of human biological monitoring (which is subject to error,
fatigue, and nighttime interruptions) with a telemetric system. It
qualitatively transforms the concept of aquaculture from prevention to
real-time action protocols, thereby laying the logical and methodological
foundations for the large-scale viability of 21st-century closed and intensive
aquatic ecosystems. The results obtained demonstrate the technical feasibility
of using computer vision and artificial intelligence to automate the behavioral
monitoring of farmed fish and generate early-warning mechanisms that contribute
to improving productivity, sustainability, and animal welfare in intensive
aquaculture systems.
References
Ahmed, M. S., & Jeba, S. M.
(2024). SalmonScan: A novel image dataset for machine
learning and deep learning analysis in fish disease detection in aquaculture. Data
in Brief, 54, 110388.
https://doi.org/10.1016/j.dib.2024.110388
Al-Abri, S., Keshvari,
S., Al-Rashdi, K., Al-Hmouz, R., & Bourdoucen, H. (2025). Computer vision-based approaches for
fish monitoring systems: a comprehensive study. Artificial Intelligence
Review, 58(6), 185. https://doi.org/10.1007/s10462-025-11180-3
Biazi, V., & Marques, C. (2023).
Industry 4.0-based smart systems in aquaculture: A comprehensive review. Aquacultural
Engineering, 103, 102360.
https://doi.org/10.1016/j.aquaeng.2023.102360
Cai, Y., Yao, Z., Jiang, H., Qin, W.,
Xiao, J., Huang, X., ... & Feng, H. (2024). Rapid detection of fish with
SVC symptoms based on machine vision combined with a NAM-YOLO v7 hybrid model. Aquaculture,
582, 740558. https://doi.org/10.1016/j.aquaculture.2024.740558
Cui, M.,
Liu, X., Liu, H., Zhao, J., Li, D., & Wang, W. (2025). Fish tracking,
counting, and behavior analysis in digital aquaculture: A comprehensive survey.
Reviews in Aquaculture, 17(1), e13001. https://doi.org/10.1111/raq.13001Digital
Object Identifier (DOI)
Fitzgerald, A., Ioannou,
C. C., Consuegra, S., Dowsey, A., & García de Leaniz, C. (2025). Machine
vision applications for welfare monitoring in aquaculture: challenges and
opportunities. Aquaculture, Fish and Fisheries, 5(1), e70036.
https://doi.org/10.1002/aff2.70036
García, L. V. (2024).
Computer vision simulator for the detection, tracking, and distance calculation
of moving objects. European Public & Social Innovation Review, 9,
1–16. https://doi.org/10.31637/epsir-2024-812
He, Q., Yu, H., Qin, H., Mei, Y., Xu,
L., Chai, Y., ... & Chen, Y. (2026). Deep learning-based computer vision
for fish behavior recognition in intensive aquaculture: A comprehensive review.
Computer Science Review, 60, 100896. https://doi.org/10.1016/j.cosrev.2026.100896
Li, X., Zhao, S., Chen, C., Cui, H.,
Li, D., & Zhao, R. (2024). YOLO-FD: An accurate fish disease detection
method based on multi-task learning. Expert Systems with Applications,
258, 125085. https://doi.org/10.1016/j.eswa.2024.125085
Liu, C., Wang, Z., Li, Y., Zhang, Z.,
Li, J., Xu, C., ... & Duan, Q. (2023). Research progress of computer vision
technology in abnormal fish detection. Aquacultural Engineering, 103,
102350. https://doi.org/10.1016/j.aquaeng.2023.102350
Ranjan, R. (2026). YOLO in Precision
Aquaculture: A Decadal Bibliometric and Systematic Review of Applications,
Architectural Adaptations, and Deployment Challenges. Journal of Agriculture
and Food Research, 102982.
https://doi.org/10.1016/j.jafr.2026.102982
Rivadeneira, F., Yi, E.
A. C., Miyahira, A., Zinanyuca, M., & Cuellar, F. (2024). Comparative evaluation of YOLO models
for gauge detection. In 2024 Latin American Robotics Symposium (LARS) (pp.
1–5). IEEE. https://doi.org/
10.1109/LARS64411.2024.10786467
Sun, Y., Liu, P., Bakarić, M. B., Yu,
J., Kong, C., & Zhang, X. (2025). Contact and non-contact physiological
stress indicators in aquatic models: A review. Aquaculture,
596, 741830. https://doi.org/10.1016/j.aquaculture.2024.741830
Terven,
J., Córdova-Esparza, D. M., & Romero-González, J. A. (2023). A comprehensive review of YOLO
architectures in computer vision: From YOLOv1 to YOLOv8 and YOLO-NAS. Machine
Learning and Knowledge Extraction, 5(4), 1680–1716. https://doi.org/10.3390/make5040083
Zhang, S., Li, D., Zhao, J., Yao, M.,
Chen, Y., Huo, Y., ... & Wang, H. (2025). Research advances on fish feeding
behavior recognition and intensity quantification methods in aquaculture. arXiv e-prints, arXiv-2502. https://doi.org/10.48550/arXiv.2502.15311