基于注意力机制的语义增强损失函数与全景分割  

Attention based semantic enhancement lossfunction and panoptic segmentation

作　　者：郑立冬[1] 滕书华 谭志国 元志安 马燕新 ZHENG Li-dong;TENG Shu-hua;TAN Zhi-guo;YUAN Zhi-an;MA Yan-xin(Qian'an Vocational Education Center,Qian'an 064400,China;College of Electronic Information,Hunan First Normal University,Changsha 410205,China;College of Meteorology and Oceanography,National University of Defense Technology,Changsha 410073,China)

机构地区：[1]河北省迁安市职业技术教育中心,河北迁安064400 [2]湖南第一师范学院电子信息学院,湖南长沙410205 [3]国防科技大学气象海洋学院,湖南长沙410073

出　　处：《激光与红外》2023年第9期1449-1456,共8页Laser & Infrared

基　　金：湖南省自然科学基金项目(No.2023JJ0185);湖南省教育厅科学研究重点项目(No.22A0640)资助。

摘　　要：全景分割是计算机视觉中重要的研究方向。考虑到不同应用场景对语义分割精度的要求不同,本文提出一种基于注意力机制的语义增强损失函数和全景分割方法。首先将语义类别按照重要程度分组,加入注意力机制来对不同语义信息进行区分,并通过对损失权重的设计有效抑制了分类失衡问题;其次设计一种全景分割网络,利用MaskR CNN网络作为实例分割子分支并加入FPN结构作为语义分割基准,提高了所需物体种类的分割精度;最后通过设计重叠结果剔除规则避免了网络结构中的实例和语义分割分支输出的重叠问题。通过对COCO数据集的对比实验表明,本文提出的语义增强损失函数有效提高了优先级较高语义类别的分割效果,为不同应用场景的全景分割提供了更加高质量的语义信息。Panoramic segmentation is an important research direction in computer vision.Considering that different application scenarios have different requirements for semantic segmentation accuracy,a semantic enhancement loss function and panoramic segmentation method based on attention mechanism is proposed in this paper.Firstly,the semantic categories are grouped according to their importance,and the attention mechanism is added to distinguish different semantic information,and the classification imbalance is effectively suppressed through the design of loss weight.Secondly,a panoramic segmentation network is designed using MaskR CNN network as the instance segmentation sub branch and adding FPN structure as the semantic segmentation benchmark to improve the segmentation accuracy of the required object types.Finally,the overlapping problem of instance and semantic segmentation branch output in network structure is avoided by designing overlapping result elimination rules.The comparative experiments on COCO data sets show that the semantic enhancement loss function proposed in this paper effectively improves the segmentation effect of semantic categories with higher priority,and provides more high quality semantic information for panoramic segmentation of different application scenarios.

关 键 词：损失函数 注意力机制 全景分割 实例分割 语义分割 

分 类 号：TN249[电子电信—物理电子学] TP391[自动化与计算机技术—计算机应用技术]