基于区域特征聚类的RGBD显著性物体检测  

RGBD Saliency Object Detection Based on Regional Feature Clustering

作　　者：王帅 蒲宝明[2] 李相泽[3] 杨朔 常战国 WANG Shuai;PU Bao-ming;LI Xiang-ze;YANG Shuo;CHANG Zhan-guo(University of Chinese Academy of Sciences,Beijing 100039,China;Shenyang Institute of Computing Technology,Chinese Academy of Sciences,Shenyang 110168 ,China;Northeastern University,School of Computer Science and Engineering,Shenyang 110819,China)

机构地区：[1]中国科学院大学,北京100039 [2]中国科学院沈阳计算技术研究所,沈阳110168 [3]东北大学计算机科学与工程学院,沈阳110819

出　　处：《小型微型计算机系统》2019年第4期704-709,共6页Journal of Chinese Computer Systems

摘　　要：本文根据在显著性检测领域的问题,提出一种基于区域特征聚类的RGBD显著性物体检测算法.首先使用结合深度信息的超像素算法对图片分割,提取分割后每个区域的特征构成特征向量.然后使用十个不同带宽的Mean Shift算法对特征向量聚类得到聚类图,并对十个聚类后的图进行显著性计算.通过神经网络把十个显著性图合并成一个显著性图,并把该显著性图作为一个新特征加到上面提到在特征向量中.继续计算显著性图,直到循环达到十次,输出最终的显著性图.通过实验,在三个RGBD显著性物体数据库中把本算法通过和七个算法进行对比,显示出本算法有更好的性能.This paper proposes an RGBD saliency object detection algorithm based on regional feature clustering based on the issues in the field of saliency detection. First,the image segmentation is performed using a superpixel algorithm that combines depth information,and the features of each region after segmentation are extracted to form feature vectors. Then ten different bandwidths of MeanShift algorithm are used to cluster the eigenvectors to obtain clustering graphs,and the ten clustered graphs are calculated for saliency values. The ten saliency maps were merged into one saliency map through a neural network,and the saliency map was added as a new feature to the eigenvector mentioned above. Continue to calculate the saliency map until the cycle reaches ten times and output the final saliency map. Through experiments,this algorithm is compared with seven algorithms in three RGBD saliency object databases,show-ing that the algorithm has better performance.

关 键 词：显著性物体检测 区域聚类 神经网络 超像素 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]