Long-term, automated fish detection provides invaluable data for deep-sea aquaculture, which is crucial for safe and efficient seawater aquafarming. In this paper, we used an infrared camera installed on a deep-sea truss-structure net cage to collect fish images, which were subsequently labeled to establish a fish dataset. Comparison experiments with our dataset based on Faster R-CNN as the basic objection detection framework were conducted to explore how different backbone networks and network improvement modules influenced fish detection performances. Furthermore, we also experimented with the effects of different learning rates, feature extraction layers, and data augmentation strategies. Our results showed that Faster R-CNN with the EfficientNetB0 backbone and FPN module was the most competitive fish detection network for our dataset, since it took a significantly shorter detection time while maintaining a high AP50 value of 0.85, compared to the best AP50 value of 0.86 being achieved by the combination of VGG16 with all improvement modules plus data augmentation. Overall, this work has verified the effectiveness of deep learning-based object detection methods and provided insights into subsequent network improvements.

This study investigated the structural response characteristics of a novel single-point mooring gravity-type deep-water (SPM-GDW) net cage under irregular waves and currents. A hydrodynamic numerical model of the cage was created and validated through model experiments. Based on the validated cage model, the structural response characteristics such as cage motion response, mooring line forces, and floating collar stress were studied, considering the actual operating conditions in the target sea area. The response time history curves, wave height time history, and spectral density statistics were studied and compared. The results showed that the heave motion of the cage was consistent with wave elevation in the vertical direction and mainly influenced by wave conditions. The surge motion of the cage was closely related to the current, with a significant lag effect compared to wave elevation motion. Low-frequency loads under the combined action of waves and currents had a significant impact on the surge motion of the cage. In addition, the mooring line tension and pontoon stress were closely related to the wave elevation, with peak values of tension and stress occurring almost simultaneously with the peak wave elevation. However, the pontoon stress exhibited high-frequency response characteristics while satisfying the wave frequency response trend. It was found that the flow velocity had a significant impact on the spectral density of mooring line tension and pontoon stress in the low-frequency range, with an increase in spectral density values as the flow velocity increased. The structural response characteristics identified in this study provide a computational basis for the optimized design and analysis of single-point mooring gravity-type deep-water cages.

Net cleaning robots have been playing an increasingly important role in offshore aquaculture due to their efficiency and labor-saving capabilities. However, in practice, these robots are still entirely teleoperated and require constant, skilled human operation. The mesh intersection points, which serve as a structural feature of the nets, provide valuable visual cues for robot self-localization and net damage identification. Therefore, the detection and tracking of these points are crucial for developing autonomous net cleaning robots. To achieve intersection point detection, we propose NPUNet-lite, a lightweight model based on U-Net. This model significantly minimizes computational resources and model size while preserving high detection accuracy. For reliable point tracking, we develop the NlPTrack algorithm, which incorporates an iterative closest point-based association strategy to meet spatial constraints between points within a frame, and a cascaded association strategy to satisfy homographic and epipolar constraints across adjacent frames. We build a dataset from videos collected during a robotic cleaning task to train and evaluate our methods. The experimental results indicate that our segmentation network achieves comparable accuracy to advanced networks, yet with a substantial reduction in computational cost. Meanwhile, the tracking method successfully tracks the majority of intersection points across scenarios where the robot moves in different directions.