DeepFins: Capturing dynamics in underwater videos for fish detection

The monitoring of fish in their natural habitat plays a crucial role in anticipating changes within marine ecosystems. Marine scientists have a preference for automated, unrestricted underwater video-based sampling due to its non-invasive nature and its ability to yield desired outcomes more rapidly compared to manual sampling. Generally, research on automated video-based detection using computer vision and machine learning has been confined to controlled environments. Additionally, these solutions encounter difficulties when applied in real-world settings characterized by substantial environmental variability, including issues like poor visibility in unregulated underwater videos, challenges in capturing fish-related visual characteristics, and background interference. In response, we propose a hybrid solution that merges YOLOv11, a popular deep learning based static object detector, with a custom designed lightweight motion-based segmentation model. This approach allows us to simultaneously capture fish dynamics and suppress background interference. The proposed model i.e., DeepFins attains 90.0% F1 Score for fish detection on the OzFish dataset (collected by the Australian Institute of Marine Science). To the best of our knowledge, these results are the most accurate yet, showing about 11% increase over the closest competitor in fish detection tasks on this demanding benchmark OzFish dataset. Moreover, DeepFins achieves an F1 Score of 83.7% on the Fish4Knowledge LifeCLEF 2015 dataset, marking an approximate 4% improvement over the baseline YOLOv11. This positions the proposed model as a highly practical solution for tasks like automated fish sampling and estimating their relative abundance.