基于视觉显著图的物体检测  

Object detection based on saliency map

作　　者：杨磊[1] 施美玲[1] 彭国琴[1] 徐丹[1] 

机构地区：[1]云南大学计算机科学与工程系,昆明650091

出　　处：《计算机应用》2010年第12期82-85,共4页journal of Computer Applications

基　　金：国家自然科学基金资助项目(60663010);高等学校博士学科点专项科研基金项目(20095301110006);云南大学研究生科研课题资助项目(ynuy200927)

摘　　要：提出了一种基于视觉显著图的显著物体检测方法。对自底向上视觉显著图计算模型进行了改进,首先对输入的彩色图片提取强度、颜色、方向三类特征金字塔;增加了轮廓弯曲度的检测,目的是使显著图形状更逼近物体;然后采用"中心—周围"算法得到三类特征图,经归一化后线性合并成最终的视觉显著图;在得到的显著图的基础上,用区域生长算法进行分割,分割的目的是为了能更显著地突出物体的大致轮廓,也减小了物体周围像素对物体的影响,分割后得到阈值化的视觉显著图。最后,对阈值化的显著图进行有效子窗口搜索,从而检测得到显著物体的位置。实验结果表明了所提方法的有效性。This paper presented a method which could detect salient object based on saliency map.This method added curvature to bottom-up visual attention model in order to obtain more approximate shape of the object.Firstly,intensity,color and orientation feature maps were extracted by bottom-up visual attention model.Total three normalized feature maps obtained by center-surround approach were linearly combined to saliency map.Curvature saliency map was processed to highlight more salient shape and reduce the negative effect of the surrounding pixels by region-grow segmentation algorithm.Salient object location was obtained by effective sub-window search in saliency map which was processed by threshold.Experimental results show the effectiveness of this method.

关 键 词：视觉显著图 区域生长分割 轮廓弯曲度 有效子窗口搜索 物体检测 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]