多分辨率线段提取方法及线段语义分析

On the Extraction of Line Segments with Multi-Resolution and Their Semantic Analysis

线段是一种组成几何体的基本元素,蕴含着非常丰富的几何信息。从图像中提取完整、连续且具有语义信息的线段对恢复场景的几何结构具有重要意义。该文提出了一种多分辨率线段提取方法,并对线段进行语义分析以区分轮廓线段和纹理线段。该方法首先运用多分辨率思想进行线段提取,然后结合深度神经网络技术对线段进行语义分析,最后对线段进行聚类合并得到最终结果。在线段连续性和完整性方面,该文提出的方法与当前常用的线段提取方法相比具有明显优势;在语义分析准确性方面,该文提出的方法在测试集上的像素精度高达97.82%。

Line segment is the essential element of geometry objects, which contains very rich geometric information. Extracting complete and continuous line segments with semantic information from an image is of great significance for restoring the geometry structure of a scene, yet challenging. This paper proposes a multiresolution segment extraction approach, which performs semantic analysis on the line segments to distinguish the contour and the texture line segments. This approach first extracts line segments with multi-resolution thought, then combines the deep neural network technology to perform semantic analysis on line segments, and finally clusters the line segments to get the final result. In terms of line segment continuity and integrity, the proposed approach has obvious advantages compared with the commonly used line segment extraction methods. In terms of semantic analysis accuracy, the pixel accuracy of the proposed approach on the test set is achieves 97.82%.