基于多尺度局部极值和边缘检测的目标性算法  

Objectness Algorithm Based on Multi-scale Local Extremum and Edge Detection

作　　者：方智文[1,2] 曹治国[1] 肖阳[1] 

机构地区：[1]华中科技大学自动化学院多谱信息技术国家级重点实验室,湖北武汉430074 [2]湖南人文科技学院能源与机电工程学院,湖南娄底417000

出　　处：《信号处理》2016年第8期911-921,共11页Journal of Signal Processing

基　　金：国家高技术研究发展计划(863计划)(2015AA015904);国家自然科学青年基金项目(61502187);娄底市科技计划项目(2015ZHJHK017);湖南人文科技学院校级青年基金项目(2015QN03);湖南省教育厅资助科研项目(14C0599;14C0597)

摘　　要：目标性作为目标检测的预处理算法,用于高效提取少量可靠的目标潜在区域,可替代针对复杂特征的多尺度滑动窗的分析方式,达到提升目标检测效率的目的。该文提出了一种基于多尺度局部极值和边缘检测的目标性算法。首先,基于原始图像的多尺度梯度特征,在不同尺度下利用均值滤波得到梯度强度的局部极值,并在原始图像上还原出初始目标潜在区域;然后,通过提取图像的边缘特征,计算初始目标潜在区域的目标性得分值;最后,对得分值进行尺度加权,并结合非极大值剔除冗余区域,最终输出少量可靠的目标潜在区域。通过PASCAL VOC和ILSVRC2014数据库的实验对比,该算法给定1000个候选区时在PASCAL VOC和ILSVRC2014分别达到97%和98%以上的召回率,同时有效地提升了首框召回率。Aiming to promote the efficiency of object detection, the objectness was introduced to pre-analyze the potential location of objects instead of the sliding window strategy with the complex features. Based on the multi-scale local extremum and edge detection, an objectness method was proposed to leverage the efficiency. First, mean filter was used to obtain the local extremum on the multi-scale gradient feature maps. According to these local extremums, coarse object proposals are extracted on the original RBG image. Second, the objectness score of each coarse object proposal was calculated based on the edge feature. Finally, redundant proposals were removed by non-maximum suppression with the scale information and the objectness score. The comparative experiment results in the public datasets （PASCAL VOC and ILSVRC 2014） demonstrated that the recall rate of our method achieved over 97% （PASCAL） and 98% （ILSVRC 2014） with 1000 proposals respectively. Furthermore, the recall rate of the top one proposal was improved too.

关 键 词：目标检测 多尺度 局部极值 边缘检测 目标性 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]