基于基本块分组与融合的特征直线检测算法-《计算机技术与发展》

文章信息/Info

Title:: A Feature Line Segment Detection Method Based on Grouping and Fusion of Primary Chunk

Author(s):: MA Rui; HU Li-hua; ZUO Wei-jian; LIU Ai-qin; School of Computer Science and Technology,Taiyuan University of Science and Technology,Taiyuan 030024, China

Keywords:: feature line segment detection; primary chunk; grouping and fusion; edge pixel detection; Helmholtz principle error culling

摘要:: 直线特征蕴含图像中重要的几何信息, 进行精确直线检测至关重要。针对场景复杂、纹理重复对象的直线检测中存在断线多、误检测率高的问题,提出一种基于基本块分组与渐进式融合的特征直线检测方法( BPC_GF) 。该方法首先采用改进的自适应 Canny 边缘检测算法检测图像边缘点的属性;其次从边缘像素点中确定瞄点,引入基本块概念,结合贪心算法生成不同类型的基本块;然后对同一类型的基本块依据相邻基本块间主方向角度偏差和空间距离约束进行分组、渐进式融合生成候选特征直线,克服了 LSD 算法中断线及 LB_LSD 算法中短线段过融合的问题;最后利用改进 Helmholtz 原理准则剔除由噪声等外界干扰形成的虚假特征直线,得到准确特征直线集。以古建筑图像为数据集进行特征直线检测,实验结果表明,与现有算法 LB_LSD 相比,该方法的精确率平均提高了 5. 43 个百分点,F-score 提高了 6. 11 个百分点。

Abstract:: Feature line segments contain important geometric information in the image,so it is quite important to detect the feature line segments accurately. In order to solve the problem of line breaking,false check in the image line extraction of scene complex and texture repetition,we present a feature line detection method based on grouping and incremental fusion of primary chunk. Firstly,the improved adaptive Canny edge detection algorithm is used to perform edge detection. Secondly,the aim point is determined from the edge pixels, introducing the primary chunk concept to generate different types of primary chunks combined with greedy algorithm. Then depending on the angle deviation and the spatial distance constraint, the primary chunks that meet the conditions are grouped and fused to gene rate candidate feature line segments,which overcomes the problems of line breaking in the LSD algorithm and short line over fusion in LB_LSD algorithm. Finally,a new Helmholtz principle is used to verify the feature line segments and eliminate the false line segments formed by external interference such as noise. With the ancient architecture image as the data set,compared with the existing image feature line detection methods of LB-LSD,the accuracy of this algorithm increases by 5. 43 percentage points on average,and the F-score by 6. 11 percentage points.