基于直线生成的卷积霍夫线段检测  

Convolutional Hough Line Segment Detection Based on Line Generation

作　　者：聂子博 曹建军 翁年凤 余旭 王孟大 NIE Zi-bo;CAO Jian-jun;WENG Nian-feng;YU Xu;WANG Meng-da(The Sixty-third Research Institute,National University of Defense Technology,Nanjing 210007,China;School of Computer Science,Nanjing University of Information Science and Technology,Nanjing 210044,China)

机构地区：[1]国防科技大学第六十三研究所数据工程研究中心,江苏南京210007 [2]南京信息工程大学计算机学院、软件学院、网络空间安全学院,江苏南京210044

出　　处：《计算机技术与发展》2024年第5期30-36,共7页Computer Technology and Development

基　　金：国防科技大学自主创新科学基金项目(23-ZZCX-KXKY-14)。

摘　　要：直线检测作为计算机视觉的上游任务,为下游包括工业视觉、遥感图像分析等任务提供支撑。直线检测的一大方向是霍夫直线检测,但现有霍夫检测基于近似原理设计计票器,直线检测准确度不高。为提高霍夫变换线段检测的准确度,利用卷积改进霍夫直线检测的计票器并提出基于直线生成的卷积霍夫线段检测方法。利用中值滤波对原始图像中的复杂纹理平滑处理后检测图像中的边界;通过按位异或卷积去除边界检测结果中的噪点并保留候选的线段端点;将候选的线段端点两两组合并使用布雷森汉姆算法进行线段生成,由利用卷积改进的计票器判断生成的线段是否存在于边界上;确认端点所构成线段位于边界后求取端点对之间的参数并合并参数相似的加检测结果,得到最终线段检测结果。对比实验中该方法的F1指标为0.7626,优于对比方法中最高的0.6523,证明该方法保留了霍夫变换较高鲁棒性的同时提高了检测结果的准确性。As an upstream task in computer vision,line detection provides support for downstream tasks including industrial vision and remote sensing image analysis.One major direction in line detection is Hough line detection.However,the accumulators in existing Hough detection methods are generally designed based on approximate principles,leading to lower accuracy in line detection.To enhance the precision of Hough line detection,convolution is employed to improve the accumulator in Hough line detection and Convolutional Hough Line Segment Detection based on line generation(CHLSD)is proposed.Firstly,median filtering is utilized to smooth the complex textures in original images,and extract edges from smoothed images.Next,the extraction results are denoised using a bitwise Exclusive OR operation and candidate end points of line segments are retained.Then,Bresenham algorithm generates line segments with paired candidate end points,accumulator improved with convolution is used to determine if generated line segments match extracted edges.Finally,parameters for determined end points are calculated and lines with similar parameters are merged to obtain final detection result.Experiments show that CHLSD improves the detection precision while retaining the robustness of Hough line detection,achieving an F1 score of 0.7626,which is superior to the comparison methods'score of 0.6523.

关 键 词：线段检测 霍夫变换 卷积核 异或 直线生成 布雷森汉姆算法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]