融合显著深度特征的RGB-D图像显著目标检测

发布时间：2019-07-21 18:21

【摘要】：深度信息被证明是人类视觉的重要组成部分,然而大部分显著性检测工作侧重于2维图像上的方法,并不能很好地利用深度进行RGB-D图像显著性检测。该文提出一种融合显著深度特征的RGB-D图像显著目标检测方法,提取基于颜色和深度显著图的综合特征,根据构图先验和背景先验的方法进行显著目标检测。首先,对原始深度图进行预处理:使用背景顶点区域、构图交点和紧密度处理深度图,多角度融合形成深度显著图,并作为显著深度特征,结合颜色特征形成综合特征;其次,从前景角度,将综合特征通过边连接权重构造关联矩阵,根据构图先验,假设多层中心矩形为前景种子,通过流形排序方法计算出RGB-D图像的前景显著图;从背景角度,根据背景先验以及边界连通性计算出背景显著图;最后,将前景显著图和背景显著图进行融合并优化得到最终显著图。实验采用RGB-D1000数据集进行显著性检测,并与4种不同的方法进行对比,所提方法的显著性检测结果更接近人工标定结果,PR(查准率-查全率)曲线显示在相同召回率下准确率高于其他方法。
[Abstract]:The depth information is proved to be an important part of human vision, but most of the significance detection work is focused on the method on the 2-dimensional image, and the RGB-D image saliency detection can not be performed well by using the depth. In this paper, a significant target detection method of RGB-D image with significant depth features is proposed, and the comprehensive features based on the color and depth saliency maps are extracted, and a significant target detection is carried out according to the method of the prior and background of the composition. firstly, pre-processing the original depth map: using a background vertex area, a composition intersection point and a compact density processing depth map, multi-angle fusion to form a depth saliency map, The method comprises the following steps of: constructing a correlation matrix through an edge connection weight; according to a composition a priori, assuming that the multi-layer center rectangle is a foreground seed, calculating a foreground saliency map of the RGB-D image by a manifold sorting method; and calculating a background saliency map from the background angle, according to the background prior and the boundary connectivity; And finally, the foreground saliency map and the background saliency map are fused and optimized to obtain the final saliency map. In the experiment, the data set of RGB-D1000 was used for the significance test, and compared with the four different methods, the significance test results of the proposed method were more close to the results of the artificial calibration, and the PR (the accuracy-checking rate) curve showed that the accuracy rate of the same recall rate was higher than that of other methods.
【作者单位】：安徽大学信息保障技术协同创新中心;安徽大学计算机科学与技术学院;
【基金】：国家科技支撑计划(2015BAK24B00) 高等学校博士学科点专项科研基金(20133401110009) 安徽高校省级自然科学研究项目(KJ2015A009) 安徽大学信息保障技术协同创新中心开放课题~~
【分类号】：TP391.41

【相似文献】