基于JETSON TX2的快速二值图像连通区域标记算法
|
摘要
为了提高二值图像连通区域标记算法速度,提出了一种基于JETSON TX2的快速二值图像连通区域标记算法。首先,创建与二值图像相同维度的标记矩阵,通过第1步扫描对标记矩阵中的非零元素设置初始化标号,再利用8邻域连通模板获取当前标号下对应邻域最小标号值;然后,通过第2步扫描寻找与判别模型相同的情况,合并等价根元素,并利用原子操作确保根元素的修改正确;最后,利用寻根操作对标记矩阵中的所有非零元素进行一次性修改。试验结果表明,该算法充分利用了GPU的并行计算能力,在处理高分辨率以及多连通域时较常见的CPU串行算法最高达到108倍的加速,与现有的GPU算法相比最高达到8倍的加速。
In order to improve the speed of connected area labeling of binary image,a fast connected area labeling algorithm on binary image based on JETSON TX2 was proposed.Firstly,a labeling matrix with the same dimension as binary image was created,the labeling was initialized for each non-zero element through the first step scan,and then the minimum labeling value of corresponding neighborhood under the current pixel was obtained by using mask for eight-connected connectivity.Then,through the second step scanning,found the same situation as the discriminant model and merged the equivalent root elements,and used atomic operation to ensure that the modification of the root elements was correct.Finally,all non-zero elements in the labeling matrix were modified one-time by root-finding operation.The experimental results showed that the algorithm made full use of the parallel computing power of GPU,when dealing with high resolution and multi-connected components,the speed of this algorithm was 108 times faster than that of the existing CPU algorithm,and was 8 times faster than that of the existing GPU algorithm.
In order to improve the speed of connected area labeling of binary image,a fast connected area labeling algorithm on binary image based on JETSON TX2 was proposed.Firstly,a labeling matrix with the same dimension as binary image was created,the labeling was initialized for each non-zero element through the first step scan,and then the minimum labeling value of corresponding neighborhood under the current pixel was obtained by using mask for eight-connected connectivity.Then,through the second step scanning,found the same situation as the discriminant model and merged the equivalent root elements,and used atomic operation to ensure that the modification of the root elements was correct.Finally,all non-zero elements in the labeling matrix were modified one-time by root-finding operation.The experimental results showed that the algorithm made full use of the parallel computing power of GPU,when dealing with high resolution and multi-connected components,the speed of this algorithm was 108 times faster than that of the existing CPU algorithm,and was 8 times faster than that of the existing GPU algorithm.
引文
[1]宋焕生,赵祥模,王养利.一种新的汽车牌照识别的图像增强算法[J].长安大学学报:自然科学版,2006,26(4):75-79.
[2]郑军,马建峰,王信义.二值图像的目标识别技术[J].机械,2002,29(2):52-55.
[3]Ronse C,Denjiver P A.Connected components in binary images:The detection problem[M].Hertfordshire:Research Studies Press,1984.
[4]Haralick R M.Some neighborhood operations[J].RealTime Parallel Computing Image Analysis,1981:11-35.
[5]赵永涛,陈庆奎,姬丽娜,等.基于CUDA的二值图像联通体标记算法[J].计算机辅助设计与图形学学报,2017,29(1):72-78.
[6]Rosenfeld A,Kak A C.Digital picture processing[M].2nd ed.New York:Academic Press,1982.
[7]He L F,Chao Y Y,Suzuki K.A linera-time two-scan labeling algorithm[C]//Image Processing,2007.ICIP 2007.IEEE International Conference on.IEEE,2007.
[8]He L F,Chao Y Y,Suzuki K,et al.Fast connectedcomponent labeling[J].Pattern Recognition,2009,42:1977-1987.
[9]Stava O,Benes B.Connected component labeling in CU-DA[M].Burlington:Morgan Kaufmann,2010.
[2]郑军,马建峰,王信义.二值图像的目标识别技术[J].机械,2002,29(2):52-55.
[3]Ronse C,Denjiver P A.Connected components in binary images:The detection problem[M].Hertfordshire:Research Studies Press,1984.
[4]Haralick R M.Some neighborhood operations[J].RealTime Parallel Computing Image Analysis,1981:11-35.
[5]赵永涛,陈庆奎,姬丽娜,等.基于CUDA的二值图像联通体标记算法[J].计算机辅助设计与图形学学报,2017,29(1):72-78.
[6]Rosenfeld A,Kak A C.Digital picture processing[M].2nd ed.New York:Academic Press,1982.
[7]He L F,Chao Y Y,Suzuki K.A linera-time two-scan labeling algorithm[C]//Image Processing,2007.ICIP 2007.IEEE International Conference on.IEEE,2007.
[8]He L F,Chao Y Y,Suzuki K,et al.Fast connectedcomponent labeling[J].Pattern Recognition,2009,42:1977-1987.
[9]Stava O,Benes B.Connected component labeling in CU-DA[M].Burlington:Morgan Kaufmann,2010.