Perception enhanced hybrid network for underwater object detection
水下目标检测技术在海洋资源勘探、环境保护等领域发挥着重要作用。然而,水下环境存在成像模糊、目标尺度多变等问题给检测任务带来了难度。为提高水下目标检测精度,提出一种感知增强混合网络。首先,构建全局-局部混合增强特征提取主干网络,利用自注意力机制求取图像中的长程全局信息,进一步构建卷积注意力增强模块提取更丰富的图像局部细节信息。通过在特征提取过程中更好地捕获图像全局和局部特征关联,增强各尺度特征描述的判别性;随后,为提高成像模糊对比低时的目标检测精度,构建目标感知增强双阶段检测头。通过增加一阶段区域提议生成网络的深度,在低质图像中提取更多目标语义信息,并通过在二阶段引入自注意力机制,抑制背景或次要信息干扰。进一步引入目标检测交并比分支,将一阶段目标先验信息融入二阶段分类求解中,提高不同种类目标的检测精度。所提方法在水下目标检测数据集TrashCan与WPBB上的<inline-formula><alternatives><math xmlns:mml="http://www.w3.org/1998/Math/MathML" id="M1"><mi>m</mi><mi>A</mi><msub><mrow><mi>P</mi></mrow><mrow><mn mathvariant="normal">0.5</mn><mo>:</mo><mn mathvariant="normal">0.95</mn></mrow></msub></math><graphic specific-use="big" xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="alternativeImage/6EAA1BED-FAC0-439a-A021-FD78DAC4E770-M001.jpg"><?fx-imagestate width="14.13933277" height="3.21733332"?></graphic><graphic specific-use="small" xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="alternativeImage/6EAA1BED-FAC0-439a-A021-FD78DAC4E770-M001c.jpg"><?fx-imagestate width="14.13933277" height="3.21733332"?></graphic></alternatives></inline-formula>和<inline-formula><alternatives><math xmlns:mml="http://www.w3.org/1998/Math/MathML" id="M2"><mi>A</mi><msub><mrow><mi>P</mi></mrow><mrow><mn mathvariant="normal">50</mn></mrow></msub></math><graphic specific-use="big" xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="alternativeImage/6EAA1BED-FAC0-439a-A021-FD78DAC4E770-M002.jpg"><?fx-imagestate width="6.60400009" height="3.21733332"?></graphic><graphic specific-use="small" xmlns:xlink="http://www.w3.org/1999/xlink" xlink:href="alternativeImage/6EAA1BED-FAC0-439a-A021-FD78DAC4E770-M002c.jpg"><?fx-imagestate width="6.60400009" height="3.21733332"?></graphic></alternatives></inline-formula>分别达到了37.8%,61.8%和82.0%,98.9%。定性和定量的对比实验结果表明,模型对不同种类水下目标均具有良好的检测准确性和鲁棒性。
Authors
- Tingting Yao (ORCID: https://orcid.org/0000-0002-9951-038X)
- Ning Li (ORCID: https://orcid.org/0000-0002-4358-6449)
- Yu Zhang (ORCID: https://orcid.org/0009-0008-0180-6555)
Publication Details
- Journal
- Optics and Precision Engineering
- Published
- 2025-01-01
- DOI
- https://doi.org/10.37188/ope.20253308.1303
- Citations
- 1
- Primary Topic
- Image Enhancement Techniques
- Type
- article
- Field-Weighted Citation Impact
- 1.28