基于优化的极限学习机和深度层次的RGB-D显著检测

RGB-D Saliency Detection Based on Optimized ELM and Depth Level

目前,相当多的显著目标检测方法均聚焦于2D的图像上,而RGB-D图像所需要的显著检测方法与单纯的2D图像相去甚远,这就需要新的适用于RGB-D的显著检测方法。该文在经典的RGB显著检测方法,即极限学习机的应用的基础上,提出融合了特征提取、前景增强、深度层次检测等多种思路的新的RGB-D显著性检测方法。该文的方法是:第一,运用特征提取的方法,提取RGB图4个超像素尺度的4096维特征;第二,依据特征提取中产生的4个尺度的超像素数量,分别提取RGB图的RGB,LAB,LBP特征以及深度图的LBE特征;第三,根据LBE和暗通道特征两种特征求出粗显著图,并在4个尺度的遍历中不断强化前景、削弱背景;第四,根据粗显著图选取前景与背景种子,放入极限学习机中进行分类,得到第1阶段显著图;第五,运用深度层次检测、图割等方法对第1阶段显著图进行再次优化,得到第2阶段显著图,即最终显著图。

Currently,many saliency-detection methods focus on 2D-image.But,these methods cannot be applied in RGB-D image.Based on this situation,new methods which are suitable for RGB-D image are needed.This paper presents a novel algorithm based on Extreme Learning Machine(ELM),feature-extraction and depth-detection.Firstly,feature-extraction is used for getting a feature,which contains 4-scale superpixels and 4096 dimensions.Secondly,according to the 4-sacle superpixels,the RGB,LAB and LBP feature of RGB image are computed,and LBE feature of depth image.Thirdly,weak salient map with LBE and dark-channel features are computed,and the foreground objects is strengthened in every circle.Fourthly,according to weak salient map,both foreground seeds and background seeds are chosen,and then,put these seeds into ELM to compute the first stage salient map.Finally,depth-detection and graph-cut are used for optimizing the first stage salient map and getting the second stage salient map.