结合渐进式特征金字塔和极化自注意力机制的海雾图像检测  

Sea fog image detection based on asymptotic feature pyramid network and polarized self-attention mechanism

作　　者：廖艺齐 熊澄丽 程钰溪 林两位 白小明[2] 李招连 LIAO Yiqi;XIONG Chengli;CHENG Yuxi;LIN Liangwei;BAI Xiaoming;LI Zhaolian(Bidding and Procurement Center of Nanchang University,Nanchang 330031,China;School of Mathematics and Computer Sciences,Nanchang University,Nanchang 330031,China;Zhangzhou Meteorological Bureau,Zhangzhou Fujian 363000,China)

机构地区：[1]南昌大学招标采购中心,江西南昌330031 [2]南昌大学数学与计算机学院,江西南昌330031 [3]福建省漳州市气象局,福建漳州363000

出　　处：《南昌大学学报（理科版）》2024年第5期490-498,共9页Journal of Nanchang University(Natural Science)

基　　金：福建省自然科学基金资助项目(2022J011079)。

摘　　要：通过对基准模型SegFormer进行两方面的改进优化,开展面向海雾图像检测方法的研究。一方面,引入渐进式特征金字塔融合模块(AFPN),有效融合海雾图像的局部和全局特征,提升模型对不同大小的海雾目标的检测能力。另一方面,引入极化自注意力机制(PSA),有效聚焦图像的细粒度空间信息,捕获海雾的边缘特征,提升模型在云雾混合区域的海雾检测能力。在真实的海雾图像数据集上进行消融实验和对比实验,所提出的模型(简称为AFPSSegFormer),与基准模型相比,mIoU、Precision和m PA指标分别提升了2.38%、2.78%与0.31%,验证了所提出模型在海雾检测方面的有效性。SegFormer as the basic model,was optimized and improved for sea fog image detection from two aspects.On the one hand,the Asymptotic Feature Pyramid Network(AFPN)was introduced to the basic model,which merged local and global features of sea fog images effectively,and improved the detection ability of the model for sea fog targets of different sizes.On the other hand,the Polarized Self-Attention(PSA)mechanism was introduced to focus on thefine-grained spatial information of images,capture the edge features of sea fog,and improve the ability of the model to detect sea fog in the mixed cloud and fog region.Ablation and comparison experiments were carried out on the real sea fog image dataset.Compared with the basic SegFormer,the indexes of mIoU,Precision and mPA were improved by 2.38%,2.78% and 0.31%respectively.The effectiveness of the proposed model in sea fog detection was verified.

关 键 词：渐进式特征金字塔 极化自注意力机制 海雾检测 SegFormer 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]