基于图Transformer的快速布局估计算法

Fast Layout Estimation Algorithm Based on Graph Transformer

在图像模式识别应用中,针对布局估计任务后处理步骤繁琐、生成算法耗时较长等问题,基于相机成像原理设计了一种快速的布局生成算法,利用图Transformer节点与邻边信息融合的特性,将复杂耗时的后处理步骤转换为平面与线段特征的交叉计算过程。使用Structured3D数据集对整体算法进行测试,结果表明各项评价指标均取得提升,单张图片计算耗时达到87 ms,整体推理速度与主流方法相比提升20%。布局生成算法耗时由单张图片66.31 ms下降至3.18 ms,速度提升了95.21%。

In the applications of image pattern recognition,a fast layout generation algorithm was designed based on camera imaging principles to address the cumbersome post-processing steps and time-consuming generation algorithms in layout estimation tasks.Additionally,leveraging the characteristics of integrating graph Transformer nodes with neighboring edge information to transform complex and time-consuming post-processing steps into a cross-computing process of plane and line segment features.The overall algorithm was tested using the Structured3D dataset,and results show that all evaluation metrics improve clearly.The computation time for a single image was reduced to 87 ms,resulting in a 20%increase in overall inference speed compared to mainstream method.The consuming time of layout generation algorithm is reduced from 66.31 ms per image to 3.18 ms,achieving a speed improvement of 95.21%.