Line segment using displacement prior

Abstract Line segment detection can offer essential technical assistance for various visual tasks, aiding computer systems in comprehending image content more effectively and executing advanced analysis and applications. Focusing on the fact that most current line segment detection methods predict the center and the displacement maps of two endpoints of a line segment separately without exploiting the coupling between them, this article proposed a dynamic label assignment strategy and an improved deformable convolution for center prediction using displacement priors, which enhances the model’s line segment sensing capability and effectively improves the detection performance. The predicted displacements are used as a priori information to guide the centroid label assignment and deformable convolution sampling of center branches, which significantly improves the performance of centroid prediction. In addition, HRNet and UNet3+ are introduced to enhance the feature expression capability of the backbone network. Finally, experiments show that the <m:math xmlns:m="http://www.w3.org/1998/Math/MathML"> <m:mi>s</m:mi> <m:mi>A</m:mi> <m:msup> <m:mrow> <m:mi>P</m:mi> </m:mrow> <m:mrow> <m:mn>10</m:mn> </m:mrow> </m:msup> </m:math> sA{P}^{10} metrics of the proposed one-stage model are 0.8 and 6.6% higher than those of the two-stage model efficient line segment detector and descriptor with the best performance in line segment detection on the Wireframe dataset and YorkUrban dataset, respectively.