Volume 2025, Issue 1 7715838

Research Article

Open Access

NanoDet Model-Based Tracking and Inspection of Net Cage Using ROV

Corresponding Author

Yinghao Wu

[email protected]

orcid.org/0000-0001-6011-3611

China North Vehicle Research Institute , No.4 Huaishuling Courtyard, Fengtai District, Beijing , 100072 , China , noveri.com.cn

Search for more papers by this author

Yaoguang Wei,

Yaoguang Wei

College of Information and Electrical Engineering , China Agricultural University , Beijing , 100083 , China , cau.edu.cn

Search for more papers by this author

Hongchao Zhang,

Hongchao Zhang

China North Vehicle Research Institute , No.4 Huaishuling Courtyard, Fengtai District, Beijing , 100072 , China , noveri.com.cn

School of Mechanical Science and Engineering , Huazhong University of Science and Technology , Wuhan , 430074 , China , hust.edu.cn

Search for more papers by this author

Yinghao Wu,

Corresponding Author

Yinghao Wu

[email protected]

orcid.org/0000-0001-6011-3611

China North Vehicle Research Institute , No.4 Huaishuling Courtyard, Fengtai District, Beijing , 100072 , China , noveri.com.cn

Search for more papers by this author

Yaoguang Wei,

Yaoguang Wei

College of Information and Electrical Engineering , China Agricultural University , Beijing , 100083 , China , cau.edu.cn

Search for more papers by this author

Hongchao Zhang,

Hongchao Zhang

China North Vehicle Research Institute , No.4 Huaishuling Courtyard, Fengtai District, Beijing , 100072 , China , noveri.com.cn

School of Mechanical Science and Engineering , Huazhong University of Science and Technology , Wuhan , 430074 , China , hust.edu.cn

Search for more papers by this author

First published: 26 February 2025

https://doi.org/10.1155/are/7715838

Academic Editor: Christyn Bailey

Share a link

Email
Wechat
Bluesky

Abstract

Open sea cage culture has become a major trend in mariculture, with strong wind resistance, wave resistance, anti-current ability, high degree of intensification, breeding density, and high yield. However, damage to the cage triggers severe economic losses; hence, to adopt effective and timely measures in minimizing economic losses, it is crucial for farmers to identify and understand the damage to the cage without delay. Presently, the damage detection of nets is mainly achieved by the underwater operation of divers, which is highly risky, inefficient, expensive, and exhibits poor real-time performance. Here, a remote-operated vehicle (ROV)-based autonomous net detection method is proposed. The system comprises two parts: the first part is sonar image target detection based on NanoDet. The sonar constantly collects data in the front and middle parts of the ROV, and the trained NanoDet model is embedded into the ROV control end, with the actual output of the angle and distance information between the ROV and net. The second part is the control part of the robot. The ROV tracks the net coat based on the angle and distance information of the target detection. In addition, when there are obstacles in front of the ROV, or it is far away from the net, the D-STAR algorithm is adopted to realize local path planning. Experimental results indicate that the NanoDet target detection exhibits an average accuracy of 77.2% and a speed of approximately 10 fps, which satisfies the requirements of ROV tracking accuracy and speed. The average tracking error of ROV inspection is less than 0.5 m. The system addresses the problem of high risk and low efficiency of the manual detection of net damage in large-scale marine cage culture and can further analyze and predict the images and videos returned from the net. https://youtu.be/NKcgPcej5sI.

1. Introduction

The world’s ocean covers an area of approximately 360 million square kilometers, with huge potential for the development and utilization of marine living resources [1, 2]. Aquatic products play an increasingly important role in human nutrition, and fish remains a prevalent commodity in the global market [3–6]. Recently, with the increasing demand for animal protein and the severe decline in conventional fishery resources, the world’s fisheries, with Marine fishing as the main body, are gradually transforming to aquaculture. Mariculture has become the primary approach to achieving a sustainable supply of aquatic products [7, 8].

Owing to the high water exchange rate and low pollutant content in the open sea area, cage farming [9] and open sea platform farming have become a new mode of fishery production, thereby providing hygienic, healthy, and high-quality seafood [10–14]. In the past two decades, the production of seawater cage culture has gradually expanded worldwide. However, the marine environment is complex and unpredictable, and several factors, such as a fish epidemic, invasion of exotic species, and destruction of net cages, are uncontrollable [15]. Broken nets can cause large numbers of fish to escape and inflict severe damage if not detected in time. According to the statistics of fish escape accidents in Norway, more than 60% of escapes were caused by the failure of farming equipment [14, 16 ], which mainly included cage wear, tears, and fish bites. Therefore, it is necessary to identify the damaged position of the net in time and repair it accordingly [17–19].

Presently, net cages are primarily inspected manually. However, this approach is conducted in a complex underwater environment with low visibility and a large net area; hence, it is time-consuming, dangerous, and inefficient. However, remote-operated vehicles (ROVs) operated underwater instead of divers exhibit flexible properties and are operated with cameras to inspect the netting suits. In addition, an ROV equipped with a camera and laser [20] can calculate the relative distance and speed of the cage. To obtain position information more accurately, the ROV system integrated the Doppler velocity log (DVL) and ultrashort baseline [21]. The relative distance, angle, and speed information of ROV relative to the net can be accurately obtained. A method to guide ROVs autonomously using aquaculture netting bars has been presented [22]. By adopting a forward-looking DVL, the netting can be approximated as a plane from four DVL beam vectors. Then, the approximate direction and position of the ROV relative to the net coat are employed as the input of the line-of-sight (LOS) guidance rule to guide the ROV to the required path and complete autonomous inspection. The feasibility of the proposed method is verified via simulation and offshore experiments. Wu et al. [23] obtained the azimuth angle between the ROV and net suit via sonar, designed an ROV net suit inspection scheme based on PID and NN hybrid drive control, realized fully autonomous net suit tracking and actual net suit inspection, and achieved verified efficiency in the complex environment of the open sea. In summary, the detection of the net in deep and open sea cages has been widely researched by experts and scholars, and a series of schemes to replace artificial underwater operations have been proposed. However, owing to the high cost of sensors or the laboratory experiment stage, they cannot be applied in actual production.

Optical cameras cannot obtain clear underwater image information owing to poor vision due to poor underwater environment and turbidity. However, sonar can provide high-quality image data under turbidity and low light conditions [24]. Based on sonar data, the angle and distance data of obstacles or tracking targets can be calculated to avoid the collision between underwater robots and obstacles and achieve safe and efficient underwater target tracking [25]. In addition, compared with optical images, sonar images have smaller files, which can reduce the pressure of data transmission and processor calculation, improve the processing speed of sonar images, and provide optimal real-time performance [26, 27]. Therefore, a tracking method based on sonar image recognition is proposed in this study. The D-STAR algorithm is adopted to realize the local path planning. Finally, the actual detection of the net is realized using the onboard camera. The sonar image is utilized to locate the position of the net suit instead of the sonar scan line, which can circumvent the interference of aquatic organisms with the system, accurately calculate the distance and angle relationship between the ROV and net suit, effectively reduce the tracking error, and realize the precise inspection of the robot.

2. Materials and Methods

2.1. Robotic Platform

Here, we present the modeling of the ROV and its sensor suite employed in this study. The underwater vehicle has a nonlinear 4-DOF model. The sensors used include sonar, inertia measurement unit, and RGB camera. The robot framework is illustrated in Figure 1.

Details are in the caption following the image — **Figure 1**
Open in figure viewer PowerPoint

Robot and net-cage frame.

2.2. Experimental Environment

The experiment was conducted in Bohai Bay, approximately 5 km away from the shore. The cage diameter, water depth, and wave height were approximately 128 m, 11 m, and 0.2–0.6 m, respectively, as illustrated in Figure 2.

Because underwater is a low-light, multidisturbance environment, the underwater vehicle will be affected by waves and currents; hence, making it difficult to control. The proper angle and distance should be maintained at all times during the inspection of the underwater net. The distance between the ROV and the net is 1 m, and the RGB camera and net are taken at a vertical angle. Set the running speed of the robot to 0.3 m/s and start shooting the net with the RGB camera at a rate of 30 fps, as shown in Figures 3 and 4.

2.3. Detection and Characterization Based on a NanoDet

The proposed method is based on sonar images. In the moving process, the underwater robot continuously scans the underwater environment to generate images. First, the generated original sonar images were enhanced. The contrast-constrained adaptive histogram equalization (CLAHE) algorithm [28, 29] was utilized. To address the problem that the local histogram equalization algorithm may lead to blocking effect and overamplified noise [30, 31], the two important steps of the CLAHE algorithm include trimming the histogram and determining the mapping relationship of each pixel via linear interpolation. The principle of histogram clipping is as follows:

The size of each subblock is m × n, and the local mapping function is as follows:

(1)

where C_DF (i) and L represent the probability cumulative distribution function of the subblock gray histogram and image gray level series, respectively. The derivative of the cumulative distribution function C_DF (i) is the normalized histogram H (i), and the slope of the local mapping function m (i) is S:

(2)

The histogram height restriction is equivalent to the slope of the local mapping function, which, in turn, limits the contrast. Next, we set the limit maximum slope of S_max, then obtain the maximum allowable histogram height as follows:

(3)

If the height of the histogram is higher than H_max, the part cutoff cannot be ignored directly. The values of the cutout parts should be evenly distributed throughout the gray range to ensure that the total area of the histogram remains the same. Another important part of CLAHE is an interpolation, which can significantly reduce the block effect. As aforementioned, the CLAHE algorithm divides the image into several subblocks of the same size, which can be divided into three groups, which include the corner, boundary, and inner regions. Pixel interpolation processing is conducted for the angle region, mapping function, and subblock pixel center pixel. For a pixel in the boundary region, its mapping function is obtained via the linear interpolation of the mapping function of the center pixel of two adjacent subblocks of the pixel.

Subsequently, the enhanced sonar image is input into the trained NanoDet model. As the output of the algorithm, the net coat is detected, the bounding box with center coordinates is detected, and the distance and angle relationship between the AUV and net cost is calculated.

The NanoDet method is a single-stage anchor-free target detection model of the FCOS type [12, 32]. It adopts ATSS for target sampling. Classification and box regression were conducted using a generalized focal loss (GFL) loss function [33].

(4)

This function can eliminate the center-ness branch of FCOS, significantly circumvent convolution on this branch, and ultimately reduce the computation overhead of the detection head. It can provide the same performance as the YOLO series, is lightweight for the three major modules of the single-stage detection model (Head, Neck, Backbone), exhibits fast target detection, and is easy to train and transplant. The overall architecture of the system based on the NanoDet model is illustrated in Figure 5.

To achieve lightweight, NanoDet employs ShuffleNetV2 1.0x as the backbone network, eliminates the last layer of the network’s convolution, and extracts 8, 16, and 32 times downsampled features into PAN for multiscale feature fusion. In the feature pyramid network (FPN) module, all convolution is discarded, and only 1X1 convolution extracted from backbone network features is retained for the alignment of feature channel dimensions. Up- and downsampling are completed via interpolation. The multiscale feature map is directly added, which reduces the computation amount of the entire feature fusion module significantly.

In the detection header module, the detection header with shared weight is employed; that is, the same set of convolution prediction detection boxes is used for the multiscale feature map out of FPN, and then a learnable scale value is adopted for each layer as the coefficient to scale the predicted boxes. To achieve lightweight, we first replaced the ordinary convolution with the depth-separable convolution and reduced the number of convolution stacks from 4 to 2. Regarding the number of channels, you can obtain the parallel speedup of most inference frameworks by compressing 256 dimensions to 96, keeping it at a multiple of 8 or 16.

Sonar images are generated at 10 fps during the operation of the underwater vehicle, which is input into the NanoDet network model after enhancement, and the target center point is calculated as follows:

(5)

where (x₀, y₀), (x₁, y₁) represent the position coordinates of the diagonal of the target detection box, while (x_i, y_i) denotes the central position of the target in the sonar image of the frame i. The angle relationship between the AUV and net is expressed as follows:

(6)

where (x_Rov, y_Rov) denotes the position of the center point of the AUV’s front end. The angle relationship between the AUV and net is expressed as follows:

(7)

2.4. Path Planning Based on the D-STAR Algorithm

During underwater inspection, once the net is encountered, it may be damaged, and the fish may escape. Therefore, the ROV should maintain a certain distance when tracking the net. Here, the D-STAR algorithm [34–36] realizes the local path planning. D-STAR is easy to calculate and can adapt to the underwater dynamic and complex environment to ensure that space collision is minimized in the inspection process of underwater vehicles, thereby reducing the safety risk caused by space collision [37, 38]. In the NanoDet detection framework, the output position of the center of the detection target is considered the target position (x_i, y_i) of D-STAR, and the position of the underwater vehicle (x_Rov, y_Rov) is considered the initial position of D-STAR. The actual path planning framework of the underwater vehicle dynamic path planning based on D-STAR is illustrated in Figure 6.

As illustrated in Figure 6a, sonar images are captured according to the target center point (x_i, y_i) and position point of the underwater vehicle (x_Rov, y_Rov), and the size of the captured image is M × N, and expressed as follows:

(8)

The sonar image matrix can be described as follows:

(9)

A map of 0,1 can be generated according to the above formula, where 0 and 1 denote no obstacle and an obstacle, respectively. In this system, threshold = 80 is set for the sonar image threshold, as illustrated in Figure 6b.

The AUV searches from the target point to the starting point, and the shortest path from the target point to each node and the actual value from the current node to the target point are stored in the storage list. The coordinate transformation is mapped to the actual environment, as illustrated in Figure 6c,d, and the coordinate transformation can be expressed as follows:

(10)

where γ is measured by the attitude sensor carried by the AUV, and R denotes the radius of the cage. The robot starts to move along the shortest path, and when moving to the next node, if the target point has not changed, no calculation is required. If it encounters an underwater creature, the robot adjusts its position and recalculates the shortest path from the target point to the initial point. Underwater environment perception is completed using a sonar sensor, as expressed in Equations (8) and (9).

3. Results

3.1. Model Performance Comparison

The model was implemented based on the deep learning framework, PyTorch. In the training process, the initial value of the learning rate was 0.0001, and the cosine annealing strategy was adopted to reduce the learning rate. The number of epochs and batch size were set to 15,000 and 64, respectively. The NanoDet model was trained over 15,000 iterations. Figure 7 presents Loss and AP (IOU = 0.5) as a function of the number of iteration steps. In the first 2000 steps, the loss decreased rapidly, then gradually slowed down, and finally stabilized at ~0.55. In addition, the AP value fluctuated between 0.74 and 0.78.

As illustrated in Figure 8, the first row presents the original sonar image, the second row illustrates the enhanced sonar image and the third row presents the detection result of NanoDet. From Figure 8, we infer that the detection accuracy was improved after image enhancement, and the tracking error of the underwater vehicle was reduced.

In addition, for real-time detection and testing of 200 random sonar images, compared with the conventional algorithm for the YOLOv5 model, the speed has a significant advantage under the condition of negligible difference in accuracy. Specific parameter comparisons are presented in Table 1.

Table 1. Comparison of NanoDet and YOLOv5s models and results of the sonar image detection.

Network	My data AP (IOU = 0.5)	Resolution	Run time 1FPS	Params	Weight size
YOLOv5s	78.5	400 X 200	110.7 ms	7.2 M	14.4 M
NanoDet	77.2	400 X 200	43.2 ms	0.95 M	7.37 M

Network	200 images recognition test (accuracy)		200 enhanced images recognition test (accuracy)

YOLOv5s	98%		98.5%
NanoDet	98%		99.5%

Network	60-s continuous frame video detection (1800 frames in total)

YOLOv5s	17 frames (not detected)
NanoDet	2 frames (not detected)

In summary, this study proposed a NanoDet-based sonar image target recognition that can realize target detection in an underwater complex environment, output real-time range, and angle data, and satisfy the requirements of ROV path tracking.

3.2. Net Tracking Experiment

In the field test in Bohai Bay, the weather was clear, the wave height was less than 0.3 m, and the underwater velocity was less than 0.2 m/s. When the ROV encounters a fish, it will avoid damaging the fish by avoiding obstacles while planning a path around the obstacle and returning to the global path, as illustrated in Figure 9.

In Figure 9, an ROV avoiding obstacles in water is shown. Experimental video: https://youtu.be/xWv3cOLtlPQ.

The attitude, tracking error, distance error, and angle error of the ROV and net were analyzed. The test was conducted on the inspection of the net at different depths of 2 m, 4 m, and 6 m, respectively, and the experimental results were analyzed.

Figure 10a presents the operating speed of the AUV. The variation of speed varies with the depth of the robot. The deeper the depth, the less the disturbance from the wind and waves on the sea surface, and the smaller the variation in the speed. Figure 10b,c presents the roll and pitch angles of the robot, respectively. The deeper the robot, the less the effect of the wind and waves on the sea surface. Figure 10d presents the change in the heading angle of the underwater vehicle during a round inspection. The deeper it is, the less it is affected by waves and currents.

As illustrated in Figure 11, the tracking position errors of the ROV at different depths during the inspection are 1.65, 1.36, and 1.24 m on average for 2, 4, and 6 m, and 2.73, 2.03, and 1.82 m at maximum. As illustrated in Figure 12, the average errors of the X and Y axis during the 2-m inspection are 0.6 m and 0.9 m, respectively, and the maximum errors are 1.23 and 1.97 m. During the 4-m inspection, the average errors of the X and Y axes were 0.69 m and 1.02 m, respectively, and the maximum errors were 1.11 and 1.59 m. The average errors of the X- and Y-axes during the 6-m inspection were 0.79 m and 1.28 m, respectively, and the maximum errors were 1.82 and 1.73 m.

As illustrated in Figure 13, the distance between the AUV and net is too large. Owing to the low underwater visibility, the image of the net was blurred, which severely affects the accuracy of the damaged parts of the net. If the distance between the underwater robot and the net is negligible, the field of vision becomes smaller, and a blind area emerges with the possibility of missing detection. In the implementation, the ideal tracking distance was set to be between 0.5 and 1.5 m. The experimental results indicate that the tracking distance of the three experiment groups is basically within the ideal range. In the experiment, which was only 2-m deep, the tracking distance reached 1.6 m owing to the influence of the waves, and the underwater vehicle quickly adjusted and returned to the ideal tracking range.

Figure 14 presents the angle relationship between the camera and net during the underwater inspection of the underwater robot. Ideally, the best effect can be achieved when the camera is perpendicular to the net (Figure 4a). The larger the angle between the camera and the net, the more difficult it is to ascertain the damage to the net. Particularly, in the underwater complex environment, when the angle error between the underwater camera and net is within ±20°, the shooting and detection of the net can be completed. The experimental results indicate that the angle error between the underwater camera and the net in the three experiment groups is less than 20°.

According to the experimental data, it can be inferred that the robot can inspect the net independently. The sea environment is complex, and different depths exert different effects on the robot. According to the three groups of experimental results, the experimental results of the 2 m depth are significantly affected by waves, especially the speed and attitude of the underwater robot. At 4 and 6 m, they are less affected by waves, and the attitude of the underwater robot is relatively stable.

4. Conclusions

Here, a deep learning model based on NanoDet was designed to detect the damage of the net in the open sea cage culture. It can detect the target in actual sonar images, return the distance and angle data between the ROV and the net, and realize the tracking of the net using the ROV. RGB shoots the net and sends the shots back to the control terminal in real-time, thereby replacing the inspection of manual underwater operations. ROV underwater autonomous net inspection exhibits high flexibility, fast swimming speed, and certain anti-interference abilities. It can shoot the net comprehensively and without dead angles. It addresses the problems of poor flexibility, poor visual field, and low efficiency in manual underwater operation. The ROV system adopted for underwater net inspections is innovative and efficient. The obtained results indicate that the proposed method can effectively detect the underwater net.

According to the experimental results, the authors put forward the following prospects: (1) further optimize the sonar image detection method, filter out the outliers in the data when the sonar obtains the original data, and improve the detection accuracy; (2) further improve the intellectualization of inspection, introduce the damage detection of the net in the RGB camera return part, automatically send out an alarm if damage is detected, and completely liberate labor requirements.

Ethics Statement

This article does not contain any studies with human participants or animals performed by any of the authors.

Consent

Informed consent was obtained from all individual participants included in the study.

Disclosure

This study belongs to the field of fishery intelligent equipment and does not involve patients and related issues. No other protected material is reproduced in this paper.

Conflicts of Interest

The authors declare no conflicts of interest.

Funding

This paper has been funded by Key Projects, but it is inconvenient to publish because the project involves secrets.

Open Research

Data Availability Statement

The data will be made available upon request.

References

1 Zhao Y.-P., Niu L.-J., Du H., and Bi C.-W., An Adaptive Method of Damage Detection for Fishing Nets Based on Image Processing Technology, Aquacultural Engineering. (2020) 90, https://doi.org/10.1016/j.aquaeng.2020.102071, 102071.
10.1016/j.aquaeng.2020.102071
Web of Science® Google Scholar
2 Shi X. W., Zhao Y. P., and Bi C. W., et al.Failure Probability of Offshore High-Density Polyethylene Net Cages Under Typhoon Waves, Aquacultural Engineering. (2022) 98, https://doi.org/10.1016/j.aquaeng.2022.102254, 102254.
10.1016/j.aquaeng.2022.102254
Web of Science® Google Scholar
3 Emerenciano M. G. C., Martinez-Cordova L. R., Martinez-Porchas M., and Miranda-Baeza A., Biofloc Technology (BFT): A Tool for Water Quality Management in Aquaculture, Water Quality, 2017, 91-109.
Google Scholar
4 Hu X., Liu Y., Zhao Z., Zhao Z., Liu J., and Zhou C., Real-Time Detection of Uneaten Feed Pellets in Underwater Images for Aquaculture Using an Improved YOLO-V4 Network[J], Computers and Electronics in Agriculture. (2021) 185, no. 8, https://doi.org/10.1016/j.compag.2021.106135.
10.1016/j.compag.2021.106135
Google Scholar
5 Rudander J., Husy T., Walree P. V., and Orten P., Coherent High-Speed Acoustic Communication Inside a Commercial Atlantic Salmon Pen, Computers and Electronics in Agriculture. (2020) 175, https://doi.org/10.1016/j.compag.2020.105540, 105540.
10.1016/j.compag.2020.105540
Web of Science® Google Scholar
6 Chen D., Wang C. M., and Zhang H., Examination of Net Volume Reduction of Gravity-Type Open-Net Fish Cages under Sea Currents, Aquacultural Engineering. (2021) 92, no. 5, https://doi.org/10.1016/j.aquaeng.2020.102128, 102128.
10.1016/j.aquaeng.2020.102128
Web of Science® Google Scholar
7 Lucas J. S. and Southgate P. C., Aquaculture: Farming Aquatic Animals and Plants-Gfsoso, 2019, Wiley-Blackwell.
Google Scholar
8 Huang X. H., Guo G. X., and Tao Q. Y., et al.Dynamic Deformation of the Floating Collar of a Net Cage under the Combined Effect of Waves and Current, Aquacultural Engineering. (2018) 83, 47–56, https://doi.org/10.1016/j.aquaeng.2018.08.002, 2-s2.0-85051970014.
10.1016/j.aquaeng.2018.08.002
Web of Science® Google Scholar
9 Ta X. and Wei Y., Research on a Dissolved Oxygen Prediction Method for Recirculating Aquaculture Systems Based on a Convolution Neural Network, Computers and Electronics in Agriculture. (2018) 145, 302–310, https://doi.org/10.1016/j.compag.2017.12.037, 2-s2.0-85041305220.
10.1016/j.compag.2017.12.037
Web of Science® Google Scholar
10 Dong S., Park S. G., Kitazawa D., Zhou J., Yoshida T., and Li Q., Model Tests and Full-Scale Sea Trials for Drag Force and Deformation of a Marine Aquaculture Net Cage, Ocean Engineering, 2021, 240.
Google Scholar
11 Zhou X., Wang J., Cao X., Fan Y., and Duan Q., Simulation of Future Dissolved Oxygen Distribution in Pond Culture Based on Sliding Window-Temporal Convolutional Network and Trend Surface Analysis, Aquacultural Engineering. (2021) 95, https://doi.org/10.1016/j.aquaeng.2021.102200, 102200.
10.1016/j.aquaeng.2021.102200
Web of Science® Google Scholar
12 Zhou Z. H., Hu Y. J., Deng X., Huang D. Q., and Lin Y., Fault Detection of Train Height Valve Based on Nanodet-Resnet101, 36th Youth Academic Annual Conference of Chinese Association of Automation (YAC), 2021, YAC.
Google Scholar
13 Zhou X. H., Wang J. P., Zhang H. X., and Duan Q. L., Application of a Hybrid Improved Sparrow Search Algorithm for the Prediction and Control of Dissolved Oxygen in the Aquaculture Industry, Applied Intelligence. (2022) .
Web of Science® Google Scholar
14 Zhao Y.-P., Lian L., Bi C.-W., and Xu Z., Digital Twin for Rapid Damage Detection of a Fixed Net Panel in the Sea, Computers and Electronics in Agriculture. (2022) 200, https://doi.org/10.1016/j.compag.2022.107247, 107247.
10.1016/j.compag.2022.107247
Web of Science® Google Scholar
15 Wei Y., Wei Q., and An D., Intelligent Monitoring and Control Technologies of Open Sea Cage Culture: A Review, Computers and Electronics in Agriculture. (2020) 169, 119, https://doi.org/10.1016/j.compag.2019.105119, 105119.
10.1016/j.compag.2019.105119
Web of Science® Google Scholar
16 Fredheim A., Current Forces on Netcage Structure, Fakultet for Ingeniørvitenskap Og Teknologi, 2005.
Google Scholar
17 Liao W., Zhang S., Wu Y., An D., and Wei Y., Research on Intelligent Damage Detection of Far-Sea Cage Based on Machine Vision and Deep Learning, Aquacultural Engineering. (2022) 96, https://doi.org/10.1016/j.aquaeng.2021.102219, 102219.
10.1016/j.aquaeng.2021.102219
Web of Science® Google Scholar
18 Akram W., Hassan T., and Toubar H., et al.Aquaculture Defects Recognition via Multi-Scale Semantic Segmentation, Expert Systems With Applications. (2024) 237, https://doi.org/10.1016/j.eswa.2023.121197, 121197.
10.1016/j.eswa.2023.121197
Web of Science® Google Scholar
19 Labra J., Zuniga M. D., and Rebolledo J., et al. Robust Automatic Net Damage Detection and Tracking on Real Aquaculture Environment Using Computer Vision, Aquacultural Engineering, 2023.
10.1016/j.aquaeng.2023.102323
Google Scholar
20 Potyagaylo S., Constantinou C. C., Georgiades G., and Loizou S. G., Asynchronous UKF-Based Localization of an Underwater Robotic Vehicle for Aquaculture Inspection Operations, OCEANS 2015 - MTS/IEEE, 2015, IEEE, 1–6, 2015https://doi.org/10.23919/OCEANS.2015.7404552.
10.23919/OCEANS.2015.7404552
Google Scholar
21 Rundtop P. and Frank K., Experimental Evaluation of Hydroacoustic Instruments for ROV Navigation Along Aquaculture Net Pens, Aquacultural Engineering. (2016) 74, 143–156, https://doi.org/10.1016/j.aquaeng.2016.08.002, 2-s2.0-84981267560.
10.1016/j.aquaeng.2016.08.002
Web of Science® Google Scholar
22 Amundsen H. B., Caharija W., and Pettersen K. Y., Autonomous ROV Inspections of Aquaculture Net Pens Using DVL, IEEE Journal of Oceanic Engineering. (2021) 47, no. 1, 1–19, https://doi.org/10.1109/JOE.2021.3105285.
10.1109/JOE.2021.3105285
Web of Science® Google Scholar
23 Wu Y. H., Liu J. C., Wei Y. G., An D., Duan Y. H., and Li W. S., Intelligent Control Method of Underwater Inspection Robot in Netcage, Aquaculture Research. (2022) 53, no. 5, 1928–1938, https://doi.org/10.1111/are.15721.
10.1111/are.15721
Web of Science® Google Scholar
24 Kasetkasem T., Tipsuwan Y., Tulsook S., Muangkasem A., Leangaramkul A., and Hoonsuwan P., A Pipeline Extraction Algorithm for Forward-Looking Sonar Images Using the Self-Organizing Map, IEEE Journal of Oceanic Engineering. (2021) 46, no. 1, 206–220, https://doi.org/10.1109/JOE.2020.2978989.
10.1109/JOE.2020.2978989
Web of Science® Google Scholar
25 Noguchi Y. and Maki T., Tracking Omnidirectional Surfaces Using a Low-Cost Autonomous Underwater Vehicle, IEEE Journal of Oceanic Engineering. (2021) 46, no. 1, 11–23, https://doi.org/10.1109/JOE.2020.2972046.
10.1109/JOE.2020.2972046
Web of Science® Google Scholar
26 Petillot Y., Ruiz I. T., and Lane D. M., Underwater Vehicle Obstacle Avoidance and Path Planning Using a Multi-Beam Forward Looking Sonar, IEEE Journal of Oceanic Engineering. (2001) 26, no. 2, 240–251, https://doi.org/10.1109/48.922790, 2-s2.0-0035301670.
10.1109/48.922790
Web of Science® Google Scholar
27 Fallon M. F., Folkesson J., McClelland H., and Leonard J. J., Relocating Underwater Features Autonomously U sing Sonar-Based SLAM, IEEE Journal of Oceanic Engineering. (2013) 38, no. 3, 500–513, https://doi.org/10.1109/JOE.2012.2235664, 2-s2.0-84880570312.
10.1109/JOE.2012.2235664
Web of Science® Google Scholar
28 Setiawan A. W., Mengko T. R., Santoso O. S., and Suksmono A. B., Color Retinal Image Enhancement Using CLAHE, ICT for Smart Society (ICISS), 2013 International Conference on IEEE, 2013, IEEE.
Google Scholar
29 Huang W. H., Yang Z., Jiao S. B., and Liu D., Research on Color Image Defogging Algorithm Based on MSR and CLAHE-Chinese Automation Congress (CAC), 2020, 7301–7306.
Google Scholar
30 Mishra A., Gupta M., and Sharma P., Enhancement of Underwater Images Using Improved CLAHE, 2018 International Conference on Advanced Computation and Telecommunication (ICACAT), 2018, ICACAT.
Google Scholar
31 Honda K., Wei K., Arai M., and Amano H., CLAHE Implementation on a Low-End FPGA Board by High-Level Synthesis, 2020 Eighth International Symposium on Computing and Networking Workshops (CANDARW), 2020, CANDARW.
Google Scholar
32 Chen G. and Duan X., Vision-Based Elderly Fall Detection Algorithm for Mobile Robot, 2021 IEEE 4th International Conference on Electronics Technology (ICET), 2021, IEEE.
Google Scholar
33 Li X., Wang W., and Wu L., et al.Generalized Focal Loss: Learning Qualified and Distributed Bounding Boxes for Dense Object Detection, Advances in Neural Information Processing Systems. (2020) 33, 21002–21012.
Google Scholar
34 Liu X., Deng R., Wang J., and Wang X., COStar: A d-Star Lite-Based Dynamic Search Algorithm for Codon Optimization, Journal of Theoretical Biology. (2014) 344, 19–30, https://doi.org/10.1016/j.jtbi.2013.11.022, 2-s2.0-84890872866.
10.1016/j.jtbi.2013.11.022
CAS PubMed Web of Science® Google Scholar
35 Zhang Z., Wu J., Dai J., and He C., A Novel Real-Time Penetration Path Planning Algorithm for Stealth Uav in 3d Complex Dynamic Environment, IEEE Access. (2020) 8, no. 99, https://doi.org/10.1109/ACCESS.2020.3007496, 122757.
10.1109/ACCESS.2020.3007496
Web of Science® Google Scholar
36 Desai P. N., Prince J. L., Wueen E. M., Cruz J. R., and Grover M. R., Entry, Descent and Landing Performance of the Mars Phoenix Lander, 2011, 48, 798–798.
Google Scholar
37 Kozynchenko A. I., Analysis of Predictive Entry Guidance for a Mars Lander under High Model Uncertainties, 2011, Elsevier Ltd..
10.1016/j.actaastro.2010.08.005
Google Scholar
38 Zhang Z., Jiang J., Wu J., and Zhu X. Z., Efficient and Optimal Penetration Path Planning for Stealth Unmanned Aerial Vehicle Using Minimal Radar Cross-Section Tactics and Modified A-Star Algorithm, 2022, ISA Transactions.
Google Scholar

All articles

NanoDet Model-Based Tracking and Inspection of Net Cage Using ROV

Abstract

1. Introduction