Answers are grounded in this paper and cite their sources.
This method improves feature extraction for underwater detection, achieving 85.7% mAP in underwater environments, indicating enhanced biological analysis.
Introduction Underwater image quality is commonly affected by problems such as insufficient illumination, extensive background noise, and target occlusion. Conventional biological detection methods suffer from the limitations of weak feature extraction, high computation, and low detection efficiency. Methods We propose an efficient and lightweight YOLO network for robots to realize high-precision underwater biological detection. Firstly, a backbone network based on hybrid dilated attention (HDA) is designed to expand the receptive field and focus on key features effectively. Secondly, a mixed aggregation star (MAS) network for the neck is constructed to enhance complex structural features and detailed textures of underwater organisms. Finally, the detection head is lightweighted using multi-scale content enhancement (MCE) modules to adaptively enhance key target channel information and suppress underwater noise. Results Compared to state-of-the-art target detection algorithms in underwater robots, our method achieves 85.7.% and 87.9% mAP@0.5 on the URPC2021 and the DUO datasets, respectively, with a model size of 5.19 M, a FLOP of 6.3 G, and a FPS of 16.54. Discussion The proposed method has excellent detection performance in underwater environments with low light, turbid water, and target occlusion.
No takes yet. Share an insight, caveat, or question.
Huang et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: