ASSISTED NET HOLE DETECTION IN FISH FARMS USING COMPUTER VISION

Visibility during inspections is often reduced by low and varying lighting, turbidity, motion blur, and biofouling. Combined with the monotony of long inspection runs, these factors make lapses in attention likely, leaving damage undetected. While computer vision approaches to net damage detection have been explored [3], prior work has largely been validated on staged or controlled data such as pools, tanks, and nets that are not representative of operational conditions. Notable exceptions [4][5] evaluate on imagery from real fish cages but rely on traditional image processing and classical machine learning rather than modern object detection, and neither is validated within routine commercial inspection operations.

Figure : Steps during ROV inspections.

Figure : Steps during ROV inspections.

This work presents a YOLOv9-based detector framework [6] that runs in real time alongside the ROV pilot, flagging candidate net damage on the live video feed as a visual redundancy layer (Figure 2). Inference runs on topside hardware rather than onboard the ROV. Detections are associated across frames using ByteTrack [7] to stabilize alerts and suppress spurious single-frame detections during live review.

The model is trained and evaluated on in-situ inspection footage provided by Njord Aqua, an industry partner performing daily ROV inspections for the fish farm industry. Video was recorded at 2560x1440 resolution and 30 fps across inspections at multiple Norwegian salmon farms. Each inspection produced raw video alongside an inspection report documenting identified damage with timestamps, which served as ground truth for annotation. Train, validation, and test splits were drawn from distinct inspections and farms to avoid data leakage; split statistics are reported in Table 1. Underwater-specific augmentations were applied to address common imaging challenges such as suspended particles, motion blur, and light scattering.

Table 1: Dataset splits.

The detector is evaluated on a held-out test set of 4,620 frames drawn from 6 inspections across 6 Norwegian salmon farms not seen during training. The test set covers variation in visibility, biofouling, and lighting representative of routine operations.

Fig. 2: Method abstract overview.

At the deployed operating point (confidence threshold 0.05), the model achieves a precision of 0.89 and recall of 0.36 on the test set, with mAP@0.5 (mean average precision at IoU 0.5) of 0.63 (Table 2). Relative to a baseline trained without underwater-specific augmentations, precision improves from 0.83 to 0.89 at comparable mAP, shifting the precision-recall balance toward fewer false positives per inspection. This means the detector is not strictly better, but better-suited for a pilot-verified workflow where false-positive rate directly affects operator trust. The optimal operating point is not obvious. A higher false-positive rate may be acceptable if it surfaces additional missed damage, since each detection is verified by the pilot.

Table 2: Test set performance at confidence threshold 0.05.

The detector runs at 46 FPS at 640×640 input on NVIDIA RTX 3500, comfortably exceeding real-time requirements for integration with the live ROV video feed. Beyond quantitative performance, the system has been deployed as an assistive tool during live commercial inspections by Njord Aqua, where it has detected net damage that pilots did not see. Still, there are cases of damage flagged by the detector and missed by pilots, and damage identified by pilots but not the detector, showing that the model is not yet reliable on its own. This is precisely the condition under which the system adds value: the detector and the pilot fail on different cases, and the combination catches more than either alone.

Acknowledgements

The work described has received funding from the Research Council of Norway. We are also grateful for the contribution of the industry partners, Njord Aqua and Måsøval.

References

[1] Ø. Jensen, T. Dempster, E. Thorstad, I. Uglem, and A. Fredheim, "Escapes of fishes from Norwegian sea-cage aquaculture: causes, consequences and prevention," Aquaculture Environment Interactions, vol. 1, no. 1, pp. 71–83, 2010.

[2] H. M. Føre and T. Thorvaldsen, "Causal analysis of escape of Atlantic salmon and rainbow trout from Norwegian fish farms during 2010–2018," Aquaculture, vol. 532, p. 736002, 2021.

[3] J. Fu, D. Liu, Y. He, and F. Cheng, "Autonomous net inspection and cleaning in sea-based fish farms: A review," Computers and Electronics in Agriculture, vol. 227, p. 109609, 2024.

[4] J. Labra, M. D. Zuniga, J. Rebolledo, M. A. Ahmed, R. Carvajal, N. Jara, and G. Carvajal, "Robust automatic net damage detection and tracking on real aquaculture environment using computer vision," Aquacultural Engineering, vol. 101, p. 102323, 2023.

[5] C. Schellewald and A. Stahl, "Irregularity detection in net pens exploiting computer vision," IFAC-PapersOnLine, vol. 55, no. 31, pp. 415–420, 2022.

[6] C.-Y. Wang and H.-Y. M. Liao, "YOLOv9: Learning what you want to learn using programmable gradient information," 2024.

[7] Y. Zhang, P. Sun, Y. Jiang, D. Yu, F. Weng, Z. Yuan, P. Luo, W. Liu, and X. Wang, "Bytetrack: Multi-object tracking by associating every detection box," 2022.