基于分支混合注意力机制的布局估计模型  

Layout Estimation Model Based on Branching Hybrid Attention Mechanism

作　　者：张靖贤 杨杰[1] 周萌萌 郭传磊 ZHANG Jing-xian;YANG Jie;ZHOU Meng-meng;GUO Chuan-lei(College of Mechanical and Electrical Engineering,Qingdao University,Qingdao 266071,China;Qingdao QCIT Technology Co.,Ltd.,Qingdao 266100,China)

机构地区：[1]青岛大学机电工程学院,青岛266071 [2]青岛联合创智科技有限公司,青岛266100

出　　处：《青岛大学学报（自然科学版）》2024年第3期68-75,共8页Journal of Qingdao University(Natural Science Edition)

基　　金：山东省自然科学基金(批准号:ZR2021MF025)资助。

摘　　要：应用场景理解算法时,现有多任务学习方法存在任务冲突。为此,提出了一种高精度布局估计模型(High Accuracy Layout Model,HALayout),并基于注意力机制设计了一种分支混合注意力模块(Branch Hybrid Attention Module,BHAM)。利用分离融合结构强化模型对不同特征的分辨能力,使算法能够更好地处理特征之间的共性和差异性,提高模型的最终计算精度;使用Structured3D数据集测试HALayout性能,并设计消融实验验证BHAM网络模块有效性。与改进之前算法相比,HALayout各项性能指标均有提升,其中IoU指标提升了2.71%,达到84.11%。In the application of scene understanding algorithms,current multi-task learning methods have the issue of task conflicts.Therefore,High Accuracy Layout Model(HALayout)and Branching Hybrid Attention Module(BHAM)based on attention mechanism were proposed to solve this problem.BHAM strengthens the model's discriminative ability towards different features by employing a structure of separation and fusion.It improves the model's ability to effectively manage both commonalities and differences among features to boost the overall computational accuracy.Performance testing of the HALayout was conducted using the Structured3D dataset,along with designed ablation experiments to validate the effectiveness of the BHAM.Compared to the previous algorithm,all performance metrics of the HALayout improved clearly,with the IoU metric increasing by 2.71%to reach 84.11%.

关 键 词：布局估计 注意力机制 计算机视觉 深度学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]