基于语义对齐与图节点交互的实例分割算法  

Instance Segmentation Algorithm Based on Semantic Alignment and Graph Node Interaction

作　　者：张敏[1] 邓洋洋 李亚军 张苗辉[1] Zhang Min;Deng Yangyang;Li Yajun;Zhang Miaohui(School of Artificial Intelligence,Henan University,Zhengzhou 450046,Henan,China)

机构地区：[1]河南大学人工智能学院,河南郑州450046

出　　处：《激光与光电子学进展》2023年第22期123-130,共8页Laser & Optoelectronics Progress

摘　　要：针对主流单阶段实例分割算法因冗余语义信息造成实例掩码缺失和泄漏的问题,提出一个基于语义对齐和图节点交互的实例分割算法。在全局掩码生成阶段,设计一个语义对齐模块,通过全局映射和高斯映射评估语义信息对全局和局部语义完整性的影响,从而对冗余语义信息进行抑制。此外,在实例掩码组装阶段,设计一个图节点交互模块。该模块通过对特征图进行图结构数据变换和图节点信息交互,提取拓扑图的空间特征,补充了掩码组装信息,进一步提高了实例掩码的准确度。实验结果表明,所提算法在MS COCO数据集上实现了38.3%的平均精度均值(mAP),与其他先进算法相比,有很强的竞争力。In order to address the issues of missing and leaking instance masks caused by redundant semantic information in mainstream single-stage instance segmentation algorithms,this paper proposes an instance segmentation algorithm based on semantic alignment and graph node interaction.In the global mask generation stage,a semantic alignment module was designed to evaluate the influence of semantic information on global and local semantic integrity through global mapping and Gaussian mapping,thereby suppressing redundant semantic information.In addition,a graph node interaction module was designed in the instance mask assembly stage that extracts spatial features of the topological graph by transforming the feature map into graph-structured data and interacting with graph node information,supplementing the mask assembly information and further improving the accuracy of the instance masks.The experimental results demonstrate that the proposed algorithm achieves a mean average accuracy(mAP)of 38.3%on the MS COCO dataset,exhibiting strong competitiveness against other state-of-the-art algorithms.

关 键 词：图像处理 实例分割 语义对齐 图节点交互 MS COCO数据集 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]