<p>Marine benthic organism detection is crucial for marine scientific research and biological resource exploitation. However, the complex underwater environment, small size, and significant scale variations of marine organisms pose significant challenges for detection algorithms. Existing deep learning-based detection algorithms often suffer from poor scale generalization and insufficient feature extraction, limiting their accuracy. To address these issues, we propose a novel detection algorithm called FOMDet (Feature Map that Optimally Matches the Object Detection Algorithm). FOMDet introduces an Optimal Feature Map Matching Paradigm based on Density Spatial Clustering (OFMPDSC) to optimize the mapping between object scales and detection head feature maps. An Optimal Feature Map Matching Strategy (OFMS) is developed, leveraging data prior information to determine the best detection head, outperforming traditional three-head architectures. Additionally, a multi-branch dynamic attention (MDA) module is proposed to enhance feature extraction capabilities by exploring the relationship between multi-branch gradient flow information and dynamic attention feature acquisition. Extensive experimental results demonstrate that FOMDet outperforms state-of-the-art algorithms on various marine benthic datasets, achieving <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="371_2025_3844_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="79" /> </InlineMediaObject> <EquationSource Format="TEX">$mAP_{0.5:0.95}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>m</mi> <mi>A</mi> <msub> <mi>P</mi> <mrow> <mn>0.5</mn> <mo>:</mo> <mn>0.95</mn> </mrow> </msub> </mrow> </math></EquationSource> </InlineEquation> scores of 42.8%, 47.0%, and 30.1% on the SUODAC, URPC, and AUDD datasets, respectively, with a reduction in computation by 11%. These findings highlight FOMDet’s potential to improve marine biological resource exploitation and conservation. The code and datasets are publicly available at <a href="https://github.com/DexterYan-cn/FOMDet.git">https://github.com/DexterYan-cn/FOMDet.git</a>, accompanied by detailed usage guides to facilitate reproducibility.</p>

Optimizing feature map matching for marine benthic organism detection

期刊论文
发表日期：

2025年3月5日

查看全文

Xinzhi Li,
Yong Liu,
Peng Yan

摘要

Marine benthic organism detection is crucial for marine scientific research and biological resource exploitation. However, the complex underwater environment, small size, and significant scale variations of marine organisms pose significant challenges for detection algorithms. Existing deep learning-based detection algorithms often suffer from poor scale generalization and insufficient feature extraction, limiting their accuracy. To address these issues, we propose a novel detection algorithm called FOMDet (Feature Map that Optimally Matches the Object Detection Algorithm). FOMDet introduces an Optimal Feature Map Matching Paradigm based on Density Spatial Clustering (OFMPDSC) to optimize the mapping between object scales and detection head feature maps. An Optimal Feature Map Matching Strategy (OFMS) is developed, leveraging data prior information to determine the best detection head, outperforming traditional three-head architectures. Additionally, a multi-branch dynamic attention (MDA) module is proposed to enhance feature extraction capabilities by exploring the relationship between multi-branch gradient flow information and dynamic attention feature acquisition. Extensive experimental results demonstrate that FOMDet outperforms state-of-the-art algorithms on various marine benthic datasets, achieving $mAP_{0.5:0.95}$ $m A P_{0.5 : 0.95}$ scores of 42.8%, 47.0%, and 30.1% on the SUODAC, URPC, and AUDD datasets, respectively, with a reduction in computation by 11%. These findings highlight FOMDet’s potential to improve marine biological resource exploitation and conservation. The code and datasets are publicly available at https://github.com/DexterYan-cn/FOMDet.git, accompanied by detailed usage guides to facilitate reproducibility.