INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue XI November 2025  
Aqua Vision: Few-Shot Learning Based Efficient Fish Identification  
in Challenging Aquatic Habitats  
Vijaya J1*, Bhomika Ratna Mandavi2, Akshat Srivastava3, Debashish Padhy4  
1Assistant Professor/ Dept DSAI, IIITNR, Raipur, Chhattisgarh, India, 493661  
2PG Student/ IIITNR, Raipur, Chhattisgarh, India, 493661  
3,4UG Student/ IIITNR, Raipur, Chhattisgarh, India, 493661  
*Corresponding Author  
DOI: https://doi.org/10.51244/IJRSI.2025.12110121  
Received: 29 November 2025; Accepted: 08 December 2025; Published: 18 December 2025  
ABSTARCT  
Aquatic ecosystems play a vital role in marine biodiversity and coastal protection, yet monitoring these  
habitats remains a significant challenge due to the scarcity of labeled data for training robust detection models.  
Traditional approaches often rely on extensive labeled datasets, which are costly and time-consuming to  
obtain, leading to a critical research gap in effective fish detection methodologies. This study introduces an  
innovative approach to fish detection by leveraging few-shot learning and pseudo-labeling techniques. We  
employ SimCLR, a contrastive learning framework, to pre-train a ResNet50-based encoder on unlabeled Deep  
Fish images, thereby extracting robust feature representations. These features are then utilized to train a Faster  
R-CNN object detection model using a limited set of labeled sea grass images. To further enhance the model’s  
performance, we incorporate pseudo-labeling, a semi-supervised learning technique that generates additional  
training data from unlabeled images based on a confidence threshold. Our methodology demonstrates  
significant improvements in fish detection accuracy. The final model achieves an average precision of 0.8167  
and recall of 0.7967, outperforming other state-of-the-art models such as YOLOv5 and RetinaNet. These  
results highlight the effectiveness of combining few-shot learning with pseudo-labeling in addressing the  
challenge of limited labeled data, paving the way for more efficient and accurate marine ecosystem  
monitoring.  
Keywords: Fish detection, Few-shot learning, Pseudo- labeling, SimCLR, Faster R-CNN, Marine ecosystem  
monitoring  
INTRODUCTION  
The sustainable protection of marine ecosystems has become globally important due to increasing  
environmental changes and human pressures. Marine habitats such as coral reefs, wetlands, sea grass  
meadows, and mangroves support biodiversity, contribute to carbon sequestration, and protect coastal regions.  
They also sustain fisheries and livelihoods, playing a key role in food security. However, these ecosystems are  
increasingly threatened by habitat degradation, pollution, and climate change.  
Fish, being highly sensitive to environmental variations, serve as important bioindicators of marine health.  
Traditional monitoring methods are labor-intensive, costly, and limited in scope, making efficient fish  
identification in challenging aquatic environments essential [1-4].  
Recent advancements in artificial intelligence (AI), and in particular deep learning techniques, have  
dramatically transformed the landscape of ecological monitoring by automating the detection and  
classification of marine species in image and video data  
[5-6]. Convolutional neural networks (CNNs)  
Page 1357  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue XI November 2025  
and emerging transformer-based models have demonstrated exceptional capabilities in visual recognition  
tasks across diverse domains including medical diagnostics, autonomous driving, and remote sensing  
[
7,8]. In the context of marine ecology, these models enable automated processing of large-scale  
underwater imagery to identify fish species and behaviors, facilitating continuous and high-resolution  
biodiversity assessments [9]. Nonetheless, the underwater domain presents unique challenges such as low  
lighting, high turbidity, occlusions caused by vegetation or other organisms, and color distortions, all of  
which significantly hamper model performance and generalizability  
[10].  
A critical limitation in deploying deep learning for underwater applications is the scarcity of well-  
annotated datasets. Labeling marine species requires expert knowledge and is time-consuming due to  
ambiguous object boundaries, camouflage, and overlapping individuals in complex habitats [11]. To  
mitigate this bottleneck, recent research has explored techniques such as self-supervised learning (SSL) to  
exploit large volumes of unlabeled data by learning useful feature representations through contrastive or  
predictive tasks [12]. Complementing SSL, few-shot learning (FSL) frameworks have been developed to  
enable models to recognize novel fish species from only a handful of labeled examples by leveraging  
metric learning or meta-learning paradigms [13]. Additionally, semi-supervised learning (Semi-SL)  
methods like Fix Match utilize pseudo-labeling strategies to further harness unlabeled images and  
improve model robustness without incurring high annotation costs [14].  
Building on these advances, we introduce AquaVision, a comprehensive detection and segmentation  
framework specifically tailored to the challenges of underwater fish identification in resource-constrained  
scenarios. AquaVision synergistically combines self-supervised representation learning via SimCLR [15],  
few-shot fine-tuning inspired by Prototypical Networks [16], and semi-supervised refinement using  
pseudo-labeling techniques [17] to maximize performance while minimizing the dependency on large  
labeled datasets. By initially learning robust and generalizable features from unlabeled underwater  
imagery, AquaVision adapts effectively to new fish species and habitats with very limited annotated data.  
The iterative semi-supervised learning process further enhances detection accuracy by leveraging  
confident predictions on unlabeled data.  
Extensive evaluation against state-of-the-art object detectors such as YOLOv5  
[
18  
]
and RetinaNet  
[19]  
demonstrates that AquaVision surpasses these baselines in precision, recall, and overall robustness,  
especially in challenging underwater conditions. This makes AquaVision a promising low-resource  
solution for scalable and automated marine biodiversity monitoring, contributing towards more effective  
conservation management and fostering the integration of AI technologies within ecological research.  
LITERATURE REVIEW  
The application of artificial intelligence, particularly deep learning, to marine visual analysis has gained  
considerable momentum in recent years. Researchers have leveraged these techniques for diverse tasks  
such as fish species classification, coral reef health assessment, detection of marine debris, and  
identification of underwater anomalies  
[20,21]. Several benchmark datasets including Fish4Knowledge  
[
22], the Underwater Robot Picking Contest (URPC) dataset [23], and DeepFish [24] have played a  
pivotal role in advancing algorithm development by providing annotated underwater images for training  
and evaluation. Despite this progress, a persistent challenge remains in translating these models to real-  
world marine environments due to factors such as domain shifts, variability in water conditions, and  
image degradation caused by factors like turbidity and lighting inconsistencies. A fundamental obstacle  
in underwater visual recognition is the limited availability of high-quality annotated data. Accurate  
labeling is often hindered by occlusions, cryptic coloration of marine organisms, and the requirement of  
expert taxonomic knowledge [25].  
Zhang et al. [26] highlighted the intensive labor and potential for human error involved in manual  
annotation processes, particularly in ecologically complex habitats such as seagrass meadows and coral  
reefs. Although conventional approaches such as data augmentation, domain adaptation, and transfer  
learning have been employed to mitigate data scarcity, their effectiveness is often habitat-specific and  
Page 1358  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue XI November 2025  
does not generalize well across diverse underwater settings.  
To address the annotation bottleneck, self-supervised learning (SSL) has emerged as a promising  
paradigm. SSL techniques like SimCLR [15], MoCo 27], and BYOL 28 allow models to learn  
[
[
]
meaningful visual representations from large collections of unlabeled images by contrasting augmented  
views of the same sample. These approaches have shown considerable promise in underwater scenarios;  
where the cost and complexity of acquiring labeled datasets pose significant barriers. By extracting  
generalizable features, SSL methods enable downstream tasks such as classification and detection with  
fewer labeled samples.  
Complementary to SSL, few-shot learning (FSL) strategies have been introduced to enable efficient  
recognition of new categories using only a handful of annotated instances. Techniques such as  
Prototypical Networks [16], Matching Networks  
[
29  
]
, and Model-Agnostic Meta-Learning (MAML) [30]  
have demonstrated strong performance on low-data classification challenges. Specifically, Lu et al. [31]  
applied metric-based few-shot learning for marine biodiversity monitoring, achieving high accuracy in  
identifying fish species from limited labeled examples. FSL approaches thus provide a practical solution  
for adapting models to novel species or environments with minimal annotation effort. Semi-supervised  
learning (Semi-SL) methods further extend model capabilities by leveraging unlabeled data alongside  
limited labeled samples. Frameworks such as FixMatch [14] and Mean Teacher [17] utilize pseudo-labeling  
and consistency regularization to iteratively improve model predictions without requiring additional manual  
labels. These techniques are particularly advantageous in marine contexts where large volumes of  
unlabeled underwater video footage exist but annotations remain scarce. Incorporating such semi-  
supervised approaches enables models to achieve enhanced robustness and accuracy in diverse  
underwater conditions.  
Traditional detectors like YOLOv5 [18] and RetinaNet [19] are fast and accurate on general datasets but  
struggle with underwater challenges such as low visibility, occlusions, and dynamic lighting. AquaVision  
overcomes these limitations by combining a robust feature extractor trained with contrastive self-supervised  
learning, refined through few-shot and semi-supervised fine-tuning. While SSL, FSL, and Semi-SL have been  
applied individually, unified frameworks are lacking. AquaVision fills this gap with an end-to-end pipeline  
that adapts to new fish species, diverse habitats, and changing conditions with minimal labeled data, enhancing  
AI-assisted marine ecosystem monitoring and conservation.  
Proposed HAR Model  
The proposed framework is specifically designed to address the unique challenges of underwater fish  
classification, particularly under low-resource conditions where annotated data is limited. Our proposed  
method leverages a multi-stage deep learning pipeline that integrates self-supervised representation  
learning, few-shot object detection, and semi-supervised training with pseudo-labels to ensure enhanced  
performance, adaptability to domain shifts, and robustness to visual noise common in underwater imagery.  
The architecture comprises four core components:  
1.Self-supervised feature extraction using SimCLR on unlabeled underwater datasets,  
2.Task-specific fine-tuning via few-shot object detection using a Faster R-CNN backbone,  
3.Semi-supervised refinement via pseudo-labeling of unlabeled instances, and  
4.Joint retraining using both true and pseudo annotations to maximize learning signal.  
An overview of the system architecture is illustrated in Figure 1, capturing the data flow and  
functional modules.  
Page 1359  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue XI November 2025  
Figure 1: Architecture of the proposed AquaVision pipeline showing sequential integration of self-supervised  
learning, few-shot detection, and pseudo-label based semi-supervised refinement.  
Self supervised Representation learning  
To address the scarcity of annotated data, we begin by learning generalizable visual features using SimCLR,  
a self-supervised contrastive learning framework. We employ a ResNet-50 encoder as the backbone and  
train it on the unlabeled DeepFish dataset. The network is trained to maximize similarity between  
different augmented views of the same image, encouraging the model to learn semantically meaningful  
representations without the need for manual labels.  
Given a batch of N images, two stochastic augmentations are applied to each sample, resulting in 2N  
transformed instances. Let  
(
zi  
,
zj  
)
represent a positive pair (i.e., two views of the same image). The  
objective is to bring them closer in the latent space while pushing apart other pairs. The loss function  
employed is the Normalized Temperature-scaled Cross Entropy (NT-Xent):  
Here, sim(·, ·) denotes cosine similarity, τ is the temperature parameter, and the denominator sums over all  
possible negatives in the batch. This stage produces a feature encoder capable of capturing rich, domain-  
relevant semantic patterns without any labeled supervision.  
Few short Fine Turning Using Faster R-CNNS  
The pretrained encoder is then integrated into a two-stage object detection frame- work, Faster R-CNN, to  
facilitate fish localization and classification in underwater imagery. We employ a limited set of annotated  
samples from the Seagrass dataset for fine-tuning. This few-shot training step enables the network to adapt its  
general representations to the target task and data distribution which is shown in Algorithm 1.  
The Faster R-CNN model outputs both bounding box coordinates and object class probabilities. The total loss  
comprises a classification term Lcls and a bounding box regression loss Lreg, balanced by a scalar hyper  
parameter λ:  
L
= L  
cls  
+ λL  
reg  
FRCNN  
(2)  
Page 1360  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue XI November 2025  
Semi Supervised Pseudo Labeling  
To address the scarcity of annotated data in our fish identification task, we incorporate a semi-supervised  
learning strategy known as pseudo-labeling. This approach allows us to utilize a large pool of unlabeled  
images by treating certain model predictions as if they were ground truth labels. Following the initial  
training phase, where the object detection model (e.g., Faster R-CNN) is fine-tuned using a limited set of  
manually labeled images, the model is then used to infer bounding boxes and class labels on the  
remaining unlabeled dataset. For each prediction, the model also outputs a confidence score indicating  
the likelihood that the prediction is accurate. To ensure that only the most reliable predictions are used  
for further training, we define a confidence threshold =0.75. Any predicted object whose associated  
confidence score exceeds this threshold is considered a high-confidence prediction. These selected  
predictions are then designated as pseudo-labels. The core idea is to enrich the training dataset by  
incorporating these pseudo- labeled samples alongside the original labeled data. This extended dataset  
enables the model to learn from a broader variety of visual examples without requiring additional manual  
annotations. Although pseudo-labels may occasionally introduce noise, the use of a high confidence  
threshold helps mitigate this risk by filtering out uncertain predictions which is shown in Algorithm 2.  
This iterative process predicting on unlabeled data, selecting high-confidence pseudo-labels, and retraining the  
model with this augmented dataset enables the model to gradually improve its generalization capability. It  
effectively leverages both labeled and unlabeled data to enhance performance, particularly in scenarios where  
labeled data is scarce or costly to obtain.  
Joint Retraining Using Few-Shot with Pseudo Label  
To leverage both the manually labeled and pseudo-labeled datasets, we retrain the detection model in a  
unified supervised learning setup. The final training set is formed by merging the few-shot labeled dataset  
with the pseudo-labeled samples generated from the previous step:  
Page 1361  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue XI November 2025  
D
train  
=
D
few  
∪
D
pseudo  
(3)  
This combined dataset enables the model to benefit from the precision of high- quality human-labeled  
examples and the diversity of machine-generated annotations. The learning objective and loss function  
remain consistent with those used during initial fine-tuning. By training on this augmented dataset, the  
model is exposed to a broader distribution of visual features and object appearances, thereby enhancing  
its ability to generalize to unseen data. This joint training paradigm effectively exploits both strong  
(ground truth) and weak (pseudo-labeled) supervision to improve detection accuracy in low-resource  
settings is shown in Algorithm 3.  
Hard Example Mining and Performance Analysis  
To further enhance performance, hard example mining is implemented by identifying false positives and  
samples with low Intersection over Union (IoU) which is shown in Algorithm 4. These challenging  
samples can be recycled for adaptive re-training or curriculum learning strategies. For comprehensive  
evaluation, the model is assessed using widely recognized object detection metrics: precision, recall, F1-  
score, and mean Average Precision (mAP) . We benchmark AquaVision against several baselines  
including YOLOv5 and RetinaNet to highlight improvements in both low-shot and domain-specific  
performance.  
Experimental setup  
This section outlines the experimental environment, datasets used, preprocessing pipeline, and the evaluation  
metrics employed to assess the performance of Aqua- Vision. The experiments were designed to evaluate the  
robustness of the system under challenging underwater conditions using few-shot and semi-supervised learning  
techniques.  
Page 1362  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue XI November 2025  
Data Set Description  
DeepFish Dataset: The unlabeled dataset consists of a large collection of 40000 underwater images captured  
across diverse tropical marine habitats, similar to the labeled portion but without accompanying annotations  
such as fish count, location, or species information. These high-resolution images (1,920 × 1,080) are sourced  
from in-situ field recordings in 20 different coastal and near shore environments in tropical Australia [32].  
The Sea grass subset: It is a focused portion of the Deep Fish dataset containing 3,255 images captured in sea  
grass-dominated marine habitats. It provides a limited set of manually annotated samples, including 16  
segmentation-labeled images, making it well-suited for few-shot detection experiments. All annotations belong  
to a single class: fish, consistent with the broader Deep Fish labeling scheme. This subset reflects the visually  
complex nature of sea grass environments, where fish often appear partially occluded or camouflaged. Overall,  
it serves as a compact yet challenging benchmark for evaluating underwater fish detection models [33].  
Data preprocessing  
All images are first resized to a fixed resolution of  
pipeline. Preprocessing also includes:  
512×512 pixels to ensure uniformity across the  
• Data Augmentation: Random cropping, horizontal flipping, color jittering, and Gaussian blur are  
applied during SimCLR pretraining to create augmented views for contrastive learning.  
• Normalization: Pixel intensities are normalized using the ImageNet mean and standard deviation to  
match pretrained model expectations.  
• Bounding Box Format Conversion: For annotation compatibility, bounding box labels are converted  
to the format required by Faster R-CNN (i.e., [x min, y min, x max, y max]).  
• Confidence Thresholding: During pseudo-labeling, detections below a confidence score of 0.75 are  
filtered out.  
This preprocessing ensures consistent data quality and enables effective transfer learning from self-  
supervised features to detection tasks.  
Training Configuration  
In this section, we discuss the complete training setup used across all stages of the AquaVision pipeline. Each  
phase of the workflow started with self-supervised SimCLR pretraining, few-shot object detector fine-tuning,  
semi-supervised pseudo-label generation, and joint retraining was configured with carefully selected  
hyperparameters to ensure stable optimization, efficient learning, and strong generalization in underwater  
environments. The following tables [1-4] summarize the key parameters, optimization choices, data  
augmentations, and thresholds employed at each step of the pipeline.  
Table 1: Self-Supervised Feature Extraction Using SimCLR  
Parameter  
Value / Setting  
Description  
Dataset  
Unlabeled DeepFish  
Source for contrastive pretraining  
Initialized for SimCLR representation learning  
Fixed size for all SimCLR image pairs  
Number of image pairs per batch  
Encoder Backbone ResNet-50  
Input Resolution  
Batch Size  
512 × 512  
32  
Page 1363  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue XI November 2025  
Epochs  
100  
Total SimCLR pretraining iterations  
Optimizer for contrastive learning  
Base LR for SimCLR  
Optimizer  
Adam  
10⁻⁴  
Learning Rate  
LR Schedule  
Weight Decay  
Temperature (τ)  
Cosine Annealing  
Smooth LR decay over 100 epochs  
Regularization to stabilize learning  
Scaling factor in NT-Xent loss  
0.0001  
0.5  
Data  
Augmentation  
Random crop, flip, color jitter, Strong transforms required for SimCLR  
Gaussian blur  
Output  
simclr_resnet50.pth  
Pretrained encoder checkpoint  
Table 2: Few-Shot Object Detection Fine-Tuning (Faster R-CNN)  
Parameter  
Dataset  
Value / Setting  
Seagrass Few-Shot  
Faster R-CNN  
Description  
Small labeled dataset for adaptation  
Two-stage detector for fish localization  
Detector Architecture  
Backbone  
ResNet-50 (SimCLR pretrained) Initialized from Stage 1  
Input Resolution  
Batch Size  
512 × 512  
Fixed resolution for detector training  
8
Per-iteration processing for fine-tuning  
Number of supervised fine-tuning iterations  
Standard for detection tasks  
Base LR for fine-tuning  
Epochs  
50  
Optimizer  
SGD (momentum = 0.9)  
10⁻⁴  
Learning Rate  
Weight Decay  
LR Schedule  
0.0001  
Regularization  
Cosine Annealing  
Smooth learning rate decay  
Confidence Threshold 0.75  
IoU Threshold (HEM) 0.5  
Minimum score for predictions  
Hard-example mining threshold  
Light augmentations to avoid label distortion  
Detector checkpoint  
Data Augmentation  
Output  
Flip, crop, color jitter  
fasterrcnn_fewshot.pth  
Table 3: Semi-Supervised Pseudo-Label Generation  
Parameter  
Value / Setting  
Description  
Used to generate pseudo detections  
Dataset  
Unlabeled DeepFish  
Page 1364  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue XI November 2025  
Detector Checkpoint  
Input Resolution  
fasterrcnn_fewshot.pth  
512 × 512  
Model from Stage 2  
Same as training for consistency  
High-confidence detections only  
Removes duplicate bounding boxes  
Avoids noisy tiny detections  
Confidence Threshold 0.75  
NMS IoU Threshold  
Small Box Removal  
Minimum Area Ratio  
0.5  
Enabled  
0.005  
Filters boxes <0.5% of image area  
Directory storing generated annotations  
Output Pseudo Labels pseudo_labels/  
Combined Dataset  
combined_labeled_pseudo/ Merged true + pseudo dataset  
Table 4: Joint Retraining Using True + Pseudo Labels  
Parameter  
Dataset  
Value / Setting  
Description  
Labeled + Pseudo-Labeled Unified training set  
Input Resolution  
Batch Size  
512 × 512  
Kept consistent across all stages  
8
Detector training batch size  
Joint training iterations  
Epochs  
50  
Optimizer  
SGD  
For stable supervised learning  
Typical for detection  
Momentum  
Learning Rate  
Weight Decay  
LR Schedule  
Sampling Strategy  
0.9  
10⁻⁴  
Base LR for retraining  
0.0001  
Regularization  
Cosine Annealing  
Labeled : Pseudo = 1 : 3  
Ensures smooth convergence  
Balances real and pseudo examples  
Down-weights noisy pseudo labels  
Can be enabled every 10 epochs  
Final trained model  
Pseudo-Loss Weight 0.5  
Pseudo Refresh  
Output  
Disabled (optional)  
aquavision_final.pth  
Evaluation Metrics  
To quantitatively assess the performance of AquaVision, we use the following standard object detection  
metrics:  
•
Precision and Recall: Measure the accuracy and completeness of fish detections. High precision  
indicates fewer false positives, while high recall indicates fewer false negatives.  
•
F1 Score: Harmonic mean of precision and recall to capture the balance between both metrics.  
Page 1365  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue XI November 2025  
These metrics provide a comprehensive view of model performance across different conditions, including  
crowded, blurry, or occluded scenes common in underwater imagery.  
RESULT ANALYSIS  
This section presents an in-depth evaluation of the AquaVision framework, focusing on both qualitative and  
quantitative analyses. The performance of our proposed approach is assessed through visual comparisons and  
numerical metrics. Specifically, we evaluate the accuracy and robustness of fish detection in challenging  
underwater environments using the Seagrass dataset. Furthermore, we benchmark our results against three well  
established object detection models: Faster R-CNN, YOLOv5, and RetinaNet.  
Qualitative Analysis  
Visual inspection plays a crucial role in understanding the behavior of object detection models,  
particularly in complex natural environments. Figures 2a and Figure 2b showcase the outputs generated  
by the proposed AquaVision model on selected images from the Seagrass dataset. The model  
demonstrates a strong ability to identify multiple fish instances even under visually challenging  
conditions, including occlusions, shadows, and dense seagrass interference.  
As depicted in Figure 2, the bounding boxes generated by the model are well- aligned with visible fish  
instances, even in scenarios where the targets are partially obscured or camouflaged. The model  
maintains high confidence in its detections, reflecting robustness against environmental noise and visual  
complexity. To further validate the model’s performance, we provide a comparison with ground truth  
annotations. Figures 3 and 4 highlight the ground truth annotations and RetinaNet outputs respectively.  
Following this, Figure5 and 6 present predictions from the used Faster R-CNN and YoloV5 for visual  
comparison: From these visual comparisons, we observe that the predicted bounding boxes exhibit strong  
spatial alignment with the ground truth labels.  
Minor discrepancies are noted in heavily occluded or low-contrast regions, which can be attributed to the  
natural complexity of underwater imagery. Nevertheless, the model demonstrates a commendable ability  
to localize and differentiate fish against complex backgrounds.  
Page 1366  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue XI November 2025  
Quantitative Evaluation  
For a more rigorous assessment, the performance of the proposed detection pipeline was evaluated on a  
test set comprising 50 images. We used standard object detection metrics such as Average Precision (AP),  
Average Recall (AR), and the F1-score to quantify model effectiveness. Table 2 summarizes the  
performance of our model in comparison with two widely used baselines, YOLOv5 and RetinaNet.  
The results indicate that the proposed framework, which combines few-shot learning with pseudo-  
labeling strategies, consistently outperforms the baseline models across all evaluation metrics. Faster R-  
CNN, when enhanced with our training methodology, achieves the highest precision and recall, indicating  
superior detection accuracy and coverage.  
The comparatively lower performance of YOLOv5 and RetinaNet highlights the advantage of our training  
strategy in limited-data regimes. These findings underscore the effectiveness of integrating semi-supervised  
learning techniques to improve model performance under real-world constraints, where acquiring labeled data  
Page 1367  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue XI November 2025  
is often resource-intensive.  
Model  
Average Precision Average Recall F1-Score  
Proposed Faster R-CNN 0.8167  
0.7967  
0.6733  
0.4467  
0.8120  
0.6667  
0.4500  
YoloV5  
0.6800  
0.4600  
RetinaNet  
CONCLUSION  
The AquaVision framework effectively showcases a novel and comprehensive approach by combining  
self-supervised learning, few-shot object detection, and semi-supervised refinement techniques to tackle  
the ongoing difficulties in accurately detecting fish within challenging underwater environments. The  
complexity of such environments characterized by factors like low visibility, occlusions, and intricate  
back- grounds poses significant obstacles for traditional detection methods. AquaVision addresses these  
challenges through an innovative training pipeline.  
Central to the framework is the use of SimCLR-based self-supervised pretraining, which enables the model  
to extract robust and transferable feature representations from a large corpus of unlabeled DeepFish  
images. This pretraining step serves as a powerful foundation, allowing the system to learn general visual  
patterns without reliance on annotated data. Subsequently, this knowledge is fine-tuned on a limited, carefully  
labeled subset of Seagrass dataset images using the Faster R-CNN architecture. This few-shot learning  
approach ensures that the model adapts effectively to the target domain with minimal labeled examples.  
To further enhance detection performance, the framework incorporates pseudo- labeling, a semi-  
supervised learning strategy that leverages unlabeled images by generating high-confidence predictions as  
additional training signals. This iterative refinement expands the effective training data and improves the  
model’s ability to generalize across diverse underwater conditions. As a result, the enhanced AquaVision  
model significantly outperforms established object detection baselines such as YOLOv5 and RetinaNet in  
terms of average precision and recall metrics, achieving a precision and recall of approximately 0.82.  
Beyond numerical metrics, qualitative evaluation reveals that AquaVision consistently localizes fish  
accurately even in visually complex scenes marked by heavy occlusion, dense seagrass, and variable  
lighting conditions. This highlights the model’s robustness and practical utility in real-world scenarios  
Declarations  
Author contribution Information  
All authors are equally contributed  
.
Competing interests  
The authors declare that they have no known competing financial interests or personal  
relationships that could have appeared to influence the work reported in this paper  
.
Funding  
The authors does not receive any Fund  
Data Availability statement  
Throughout the research used the public available data set which is cited in 32 &33.  
Page 1368  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue XI November 2025  
REFERENCES  
1. Ahmed, R., & Tamim, M. T. R. (2025). Marine and Coastal Environments: Challenges, Impacts,  
and Strategies for a Sustainable Future. International Journal of Science Education and  
Science, 2(1), 53-60.  
2. Douglas, J., Niner, H., & Garrard, S. (2024). Impacts of marine plastic pollution on seagrass  
meadows and ecosystem services in Southeast Asia. Journal of Marine Science and  
Engineering, 12(12), 2314.  
3. Schmid, B., & Schöb, C. (2022). Biodiversity and ecosystem services in managed ecosystems.  
In The ecological and societal consequences of biodiversity loss (pp. 213-231). ISTE Ltd and John  
Wiley & Sons, Inc, London.  
4. Hong, J. H., Semprucci, F., Jeong, R., Kim, K., Lee, S., Jeon, D., ... & Lee, W. (2020).  
Meiobenthic nematodes in the assessment of the relative impact of human activities on coastal  
marine ecosystem. Environmental Monitoring and Assessment, 192(2), 81..  
5. Lopez-Vazquez, V., Lopez-Guede, J. M., Marini, S., Fanelli, E., Johnsen, E., & Aguzzi, J. (2020).  
Video image enhancement and machine learning pipeline for underwater animal detection and  
classification at cabled observatories. Sensors, 20(3), 726.  
6. Jalal, A., Salman, A., Mian, A., Shortis, M., & Shafait, F. (2020). Fish detection and species  
classification  
in  
underwater  
environments  
using  
deep  
learning  
with  
temporal  
information. Ecological Informatics, 57, 101088.  
7. Meena, T., Vijaya, J., & Harsha, B. (2025, February). Swin Transformers for Remote Sensing SAR  
Image Classification. In 2025 IEEE International Conference on Emerging Technologies and  
Applications (MPSec ICETA) (pp. 1-6). IEEE.  
8. Vijaya, J., Gopu, A., Suman, P., & Chaitanya, S. (2024, May). Revolutionising Image  
Enhancement Leveraging Power OF CNN’S. In 2024 3rd International Conference on Artificial  
Intelligence For Internet of Things (AIIoT) (pp. 1-6). IEEE.  
9. Yassir, A., Andaloussi, S. J., Ouchetto, O., Mamza, K., & Serghini, M. (2023). Acoustic fish  
species identification using deep learning and machine learning algorithms: A systematic  
review. Fisheries Research, 266, 106790.  
10. Fu, C., Liu, R., Fan, X., Chen, P., Fu, H., Yuan, W., ... & Luo, Z. (2023). Rethinking general  
underwater object detection: Datasets, challenges, and solutions. Neurocomputing, 517, 243-256.  
11. Er, M. J., Chen, J., Zhang, Y., & Gao, W. (2023). Research challenges, recent advances, and  
popular  
datasets  
in  
deep  
learning-based  
underwater  
marine  
object  
detection:  
A
review. Sensors, 23(4), 1990.  
12. Li, J., Yang, W., Qiao, S., Gu, Z., Zheng, B., & Zheng, H. (2024). Self-supervised marine  
organism detection from underwater images. IEEE Journal of Oceanic Engineering.  
13. Chungath, T. T., Nambiar, A. M., & Mittal, A. (2023). Transfer learning and few-shot learning  
based deep neural network models for underwater sonar image classification with a few  
samples. IEEE Journal of Oceanic Engineering, 49(1), 294-310.  
14. Sohn, K., Berthelot, D., Carlini, N., Zhang, Z., Zhang, H., Raffel, C. A., ... & Li, C. L. (2020).  
Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in  
neural information processing systems, 33, 596-608.  
15. Chen, T., Kornblith, S., Norouzi, M., & Hinton, G. (2020, November). A simple framework for  
contrastive learning of visual representations. In International conference on machine learning (pp.  
1597-1607). PmLR.  
16. Snell, J., Swersky, K., & Zemel, R. (2017). Prototypical networks for few-shot learning. Advances  
in neural information processing systems, 30.  
17. Tarvainen, A., & Valpola, H. (2017). Mean teachers are better role models: Weight-averaged  
consistency targets improve semi-supervised deep learning results. Advances in neural information  
processing systems, 30.  
18. Li, L., Shi, G., & Jiang, T. (2023). Fish detection method based on improved  
YOLOv5. Aquaculture International, 31(5), 2513-2530.  
19. Shen, Z., & Nguyen, C. (2020, November). Temporal 3D RetinaNet for fish detection. In 2020  
Digital Image Computing: Techniques and Applications (DICTA) (pp. 1-5). IEEE.  
Page 1369  
www.rsisinternational.org  
INTERNATIONAL JOURNAL OF RESEARCH AND SCIENTIFIC INNOVATION (IJRSI)  
ISSN No. 2321-2705 | DOI: 10.51244/IJRSI |Volume XII Issue XI November 2025  
20. Vyshnav, K., Sooryanarayanan, R., & Madhav, T. V. (2024, April). Analysis of Underwater Coral  
Reef Health Using Neural Networks. In OCEANS 2024-Singapore (pp. 01-06). IEEE.  
21. Chowdhury, A., Jahan, M., Kaisar, S., Khoda, M. E., Rajin, S. A. K., & Naha, R. (2024). Coral  
Reef Surveillance with Machine Learning: A Review of Datasets, Techniques, and  
Challenges. Electronics, 13(24), 5027.  
22. https://homepages.inf.ed.ac.uk/rbf/fish4knowledge/  
23. https://www.kaggle.com/datasets/lywang777/urpc2020  
24. https://datasetninja.com/deep-fish  
25. Elmezain, M., Saoud, L. S., Sultan, A., Heshmat, M., Seneviratne, L., & Hussain, I. (2025).  
Advancing underwater vision: a survey of deep learning models for underwater object recognition  
and tracking. IEEE Access.  
26. Zhang, F., Hu, J., & Sun, Y. (2025). Underwater fish image recognition based on knowledge  
graphs and semi-supervised learning feature enhancement. Scientific Reports.  
27. Dalla Serra, F., Jacenków, G., Deligianni, F., Dalton, J., & O’Neil, A. Q. (2022, July). Improving  
Image Representations via MoCo Pre-training for Multimodal CXR Classification. In Annual  
Conference on Medical Image Understanding and Analysis (pp. 623-635). Cham: Springer  
International Publishing.  
28. Niizumi, D., Takeuchi, D., Ohishi, Y., Harada, N., & Kashino, K. (2021, July). Byol for audio:  
Self-supervised learning for general-purpose audio representation. In 2021 International Joint  
Conference on Neural Networks (IJCNN) (pp. 1-8). IEEE.  
29. Berthelot, D., Carlini, N., Goodfellow, I., Papernot, N., Oliver, A., & Raffel, C. A. (2019).  
Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information  
processing systems, 32.  
30. Fallah, A., Mokhtari, A., & Ozdaglar, A. (2020, June). On the convergence theory of gradient-  
based model-agnostic meta-learning algorithms. In International Conference on Artificial  
Intelligence and Statistics (pp. 1082-1092). PMLR.  
31. Lu, J., Zhang, S., Zhao, S., Li, D., & Zhao, R. (2024). A metric-based few-shot learning method for  
fish species identification with limited samples. Animals, 14(5), 755.  
32. https://alzayats.github.io/DeepFish/  
33. https://github.com/globalwetlands/luderick-seagrass  
Page 1370  
www.rsisinternational.org