Fish-Finder: A robust small target detection method for aquaculture fish in low-quality underwater images

Area of Science:

• Computer Vision
• Marine Biology
• Aquaculture Management

Background:

• Underwater fish object detection is crucial but challenging due to environmental complexity, occlusions, and small, mobile fish.
• Existing methods struggle with the intricacies of underwater scenes, impacting accuracy in marine research and aquaculture.

Purpose of the Study:

• To propose a novel algorithm, Fish-Finder, for robust underwater fish object detection.
• To address the challenges of underwater environments and improve detection accuracy for small, moving fish.

Main Methods:

• Introduced 'C2fBF' structure with BiFormer's dual-path routing attention to preserve features during downsampling.
• Integrated RepGFPN in the neck network for effective merging of semantic and spatial information, enhancing multi-scale detection.
• Incorporated Wasserstein loss with CIoU to reduce sensitivity to positional errors in detecting small fish.

Main Results:

• The Fish-Finder algorithm demonstrated improved accuracy and mean average precision (mAP) on both public (Kaggle-Fish) and the new SmallFish datasets.
• Significant performance gains were observed compared to state-of-the-art detection methods.

Conclusions:

• Fish-Finder offers a promising solution for accurate underwater fish object detection in challenging conditions.
• The developed algorithm advances capabilities in marine biology, aquaculture, and computer vision applications.