A New Workflow for Instance Segmentation of Fish with YOLO

Jiushuang Zhang,Yong Wang

DOI: https://doi.org/10.3390/jmse12061010

IF: 2.744

2024-06-19

Journal of Marine Science and Engineering

Abstract:The application of deep-learning technology for marine fishery resource investigation is still in its infancy stage. In this study, we applied YOLOv5 and YOLOv8 methods to identify and segment fish in the seabed. Our results show that both methods could achieve superior performance in the segmentation task of the DeepFish dataset. We also expanded the labeling of specific fish species classification tags on the basis of the original semantic segmentation dataset of DeepFish and completed the multi-class instance segmentation task of fish based on the newly labeled tags. Based on the above two achievements, we propose a general and flexible self-iterative fish identification and segmentation standard workflow that can effectively improve the efficiency of fish surveys.

oceanography,engineering, marine, ocean

What problem does this paper attempt to address?

The paper primarily addresses the issue of fish image recognition and segmentation in marine fishery resource surveys, particularly improving the efficiency of fish surveys through computer vision technology. Specifically, the study employs two methods, YOLOv5 and YOLOv8, for fish instance segmentation and proposes a general and flexible workflow for fish recognition and segmentation based on these methods.

Key points of the paper include:

1. **Application Background**: Although the application of computer vision technology in fishery resource surveys is still in its early stages, the automatic identification and statistics of deep-sea fish are of great significance for ecological balance and the fishery industry with the growth of monitoring data and the improvement of computing power.

2. **Method Introduction**:
- Using YOLOv5 and YOLOv8 to identify and segment seabed fish.
- Extending the annotation work of specific fish species classification labels on the DeepFish dataset, completing the multi-class instance segmentation task based on the new annotation labels.
- Proposing a self-iterative standard workflow for fish recognition and segmentation.

3. **Experimental Results**:
- In the binary semantic segmentation task, YOLOv5 and YOLOv8 achieved mask-level mAP@[0.5, 0.95] of 0.8431 and 0.8577, respectively.
- For the multi-class instance segmentation task, on a dataset containing 19 valid fish species, YOLOv8 achieved a mask-level mAP@[0.5, 0.95] of 0.9381.
- The trained models demonstrated good generalization ability in unseen sample tests (i.e., previously collected deep-sea videos).

In summary, this study not only validates the effectiveness of YOLOv5 and YOLOv8 in fish segmentation tasks but also proposes a practical workflow that can significantly improve the efficiency of marine fishery resource surveys.