Table 1.

Summary of related work.

Authors	Title	Methodology	Merits	Demerits
Li et al. [26]	Tackling Fish Detection with Occlusion Using Modified YOLOv8	Employed Real-time Detection Transformer (RT-DETR) modules and repulsion loss for occluded fish detection.	Improved mAP compared to original YOLOv8.	Potential complexity in implementing RT-DETR and repulsion loss.
Kuswantori et al. [27]	Optimizing YOLOv4 for Fish Recognition	Utilized special labeling technique to optimize YOLOv4 recognition algorithm.	High accuracy (98.15 %) in recognizing fish in erratic arrangements.	Specific to YOLOv4, may not generalize to other algorithms.
Wang et al. [28]	DFYOLO: Enhanced Fish Disease Diagnosis Model Using YOLOv5	Enhanced YOLOv5 with CSPNet structure, attention module, and parallel convolutional kernels.	Significant increase in average precision (4.86 %) for identifying diseased fish.	Modifications may require expertise and computational resources.
Ranjan et al. [29]	Factors Affecting Precision of Fish Detection in RAS Production Settings	Investigated impact of sensor choice, dataset size, imaging parameters, and pre-processing on model precision.	Highlighted importance of sensor selection and dataset size for accurate fish detection in RAS.	Specific to RAS production settings, findings may not generalize to other scenarios.
Fernandez Garcia et al. [30]	Novel Approach for Fish Passage Detection in Acoustic Video Streams	Combined classical CV techniques with CNN for fish passage detection in acoustic video streams.	Improved performance in fish passage detection after pre-treatment of acoustic images.	Requires pre-treatment of acoustic images, potential computational overhead.
Patro et al. [31]	Real-time Fish Identification in Underwater Environments Using YOLOv5-CNN	Implemented YOLOv5-CNN model for real-time fish identification in various underwater conditions.	Achieved high accuracy (mAP: 0.86) in real-time fish detection underwater.	Specific to YOLOv5-CNN, may not generalize to other detection models.
Kandimalla et al. [32]	OceanFish: Database and Testbed for Fish Recognition	Introduced OceanFish database and testbed utilising deep neural network-based object detection models.	Reliable performance in fish recognition with improved image quality and resolution.	Limited to fish recognition, may require additional annotation efforts for new species.
Kandimalla et al. [33]	Real-time Automated Deep Learning Framework for Fish Classification	Developed real-time automated framework combining Kalman filters and CNN for fish classification.	Successful application in sonar imaging and underwater video for species classification.	Framework complexity, potential computational demands.
Al Muksit et al. [34]	YOLO-Fish: Deep Learning Model for Fish Identification in Marine Environments	Proposed YOLO-Fish models (YOLO-Fish-1 and YOLO-Fish-2) for fish identification in marine environments.	Achieved average precision scores for fish detection in real-life marine environments.	Performance may vary depending on environmental conditions and fish species diversity.
Alaba et al. [35]	Species Recognition for Fish Object Detection Using MobileNetv3-large and VGG16 Backbone Networks	Developed species recognition model using MobileNetv3-large and VGG16 backbone networks with SSD detection head.	Addressed class imbalance in dataset with a class-aware loss function.	Specific to MobileNetv3-large and VGG16, may require adaptation for other networks.
Hong Khai et al. [36]	Enhanced Mask R–CNN for Shrimp Detection on Shrimp Farm	Enhanced Mask R–CNN model for shrimp detection on shrimp farm using parameter calibration.	Achieved high accuracy (97.48 %) in identifying shrimp density categories.	Parameter calibration process may be labor-intensive, specific to shrimp detection.
Yin et al. [37]	Individual Fish Identification Using Coarse- and Fine-grained Features	Developed technique using coarse- and fine-grained feature learning networks for individual fish identification.	Achieved superior performance in fish recognition with coarse- and fine-grained features.	Complexity in feature engineering and model training, potential computational demands.